WO2020228570A1 - 乳腺钼靶图像处理方法、装置、系统及介质 - Google Patents

乳腺钼靶图像处理方法、装置、系统及介质 Download PDF

Info

Publication number
WO2020228570A1
WO2020228570A1 PCT/CN2020/088759 CN2020088759W WO2020228570A1 WO 2020228570 A1 WO2020228570 A1 WO 2020228570A1 CN 2020088759 W CN2020088759 W CN 2020088759W WO 2020228570 A1 WO2020228570 A1 WO 2020228570A1
Authority
WO
WIPO (PCT)
Prior art keywords
mammography
target image
model
image
detection
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2020/088759
Other languages
English (en)
French (fr)
Inventor
田宽
江铖
颜克洲
沈荣波
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Publication of WO2020228570A1 publication Critical patent/WO2020228570A1/zh
Priority to US17/367,266 priority Critical patent/US11922654B2/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • A61B6/12—Arrangements for detecting or locating foreign bodies
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H15/00—ICT specially adapted for medical reports, e.g. generation or transmission thereof
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/20—ICT specially adapted for the handling or processing of medical images for handling medical images, e.g. DICOM, HL7 or PACS
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • A61B6/50—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment specially adapted for specific body parts; specially adapted for specific clinical applications
    • A61B6/502—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment specially adapted for specific body parts; specially adapted for specific clinical applications for diagnosis of breast, i.e. mammography
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00—Apparatus or devices for radiation diagnosis; Apparatus or devices for radiation diagnosis combined with radiation therapy equipment
    • A61B6/52—Devices using data or image processing specially adapted for radiation diagnosis
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10116—X-ray image
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30068—Mammography; Breast
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30096—Tumor; Lesion
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images

Definitions

  • the embodiments of the present application relate to the field of artificial intelligence, and in particular to a method, device, system, and medium for processing mammary gland images.
  • Mammograms are widely used in early breast cancer screening. Doctors usually diagnose various abnormal information in mammograms, including tumors, calcifications, lymphadenopathy and skin abnormalities, etc., and based on the abnormal information give the breast imaging report and data system (BI-RADS) classification that characterizes the degree of risk.
  • BI-RADS breast imaging report and data system
  • a neural network model is used to identify abnormal areas in mammography images to locate abnormal areas in mammography images and obtain corresponding abnormal information, so as to infer all abnormal information in mammography images
  • the BI-RADS score of the whole mammography target image is obtained.
  • the prediction process of the above-mentioned neural network model is only a simple two-dimensional image-level prediction process, which does not coincide with the doctor's actual viewing process, resulting in a low degree of accuracy in identifying abnormal areas.
  • Various embodiments of the present application provide an auxiliary diagnosis method, device, computer equipment, system and medium for mammography mammography image, and a method, apparatus, computer equipment, system and medium for processing mammography mammogram image.
  • a method for assisting diagnosis of mammography target images which is executed by computer equipment, and includes:
  • a mammogram image of a unilateral breast including: a head and tail (Cranial-Caudal, CC) mammography image and a medial strabismus (MedioLateral-Oblique, MLO) mammography image;
  • test report including the benign and malignant prediction results of the unilateral breast.
  • An auxiliary diagnosis device for mammography target images comprising:
  • An image acquisition module for acquiring mammography images of a unilateral breast the mammography images including: CC mammography images and MLO mammography images;
  • the automatic report output module is used to generate and output a test report, the test report including the benign and malignant prediction results of the unilateral breast.
  • a computer device includes a memory and a processor.
  • the memory stores computer readable instructions.
  • the processor executes the computer readable instructions, the steps of the method for assisting mammography of mammography as described above are realized.
  • a mammography assisted diagnosis system for mammography comprising: breast DR equipment, computer equipment and doctor equipment; said breast DR equipment is connected to said computer equipment, and said computer equipment is connected to said doctor equipment;
  • the computer device includes a memory and a processor; the memory stores computer readable instructions, and the computer readable instructions are loaded and executed by the processor to implement the aforementioned method for assisting diagnosis of mammography.
  • a computer-readable storage medium in which computer-readable instructions are stored, and the computer-readable instructions are loaded and executed by a processor to realize the steps of the method for assisting diagnosis of mammography images as described above.
  • a mammography target image processing method executed by a computer device, the method comprising:
  • Acquire mammary mammography images of a unilateral breast the mammography images include: head and tail CC mammography images and medial oblique MLO mammography images;
  • a detection report is generated and output.
  • a mammography target image processing device comprising:
  • the image acquisition module is used to acquire mammography mammography images of a unilateral breast, the mammography mammography images including: head and tail CC mammography images and medial oblique MLO mammography images;
  • the target detection model is used to process the CC mammography target image and the MLO mammography target image to obtain the image detection result of the unilateral breast;
  • the automatic report output module is used to generate and output a detection report according to the image detection result.
  • a computer device includes a memory and a processor, the memory stores computer readable instructions, and the processor implements the steps of the mammography image processing method when the computer readable instructions are executed.
  • a mammography target image processing system comprising: a breast DR device, a computer device, and a doctor device; the breast DR device is connected to the computer device, and the computer device is connected to the doctor device; the computer
  • the device includes a memory and a processor, and the memory stores computer readable instructions, and the processor implements the steps of the mammography image processing method when the computer readable instructions are executed.
  • a computer-readable storage medium storing computer-readable instructions, which is characterized in that, when the computer-readable instructions are executed by a processor, the steps of the mammography image processing method described above are realized.
  • FIG. 1 is a flowchart of an auxiliary detection method for mammography target images provided by an exemplary embodiment of the present application
  • FIG. 2 is a flowchart of an auxiliary detection method of mammography target images provided by another exemplary embodiment of the present application
  • FIG. 3 is a structural block diagram of a breast benign and malignant model provided by an exemplary embodiment of the present application.
  • FIG. 4 is a flowchart of an auxiliary detection method for mammography target images provided by another exemplary embodiment of the present application.
  • FIG. 5 is a flowchart of a breast information extraction subsystem provided by another exemplary embodiment of the present application.
  • Fig. 6 is a flowchart of a training method for a nipple and muscle detection model provided by another exemplary embodiment of the present application;
  • FIG. 7 is a flowchart of a method for training a gland type classification model provided by another exemplary embodiment of the present application.
  • FIG. 8 is a working principle diagram of a lesion detection subsystem provided by another exemplary embodiment of the present application.
  • FIG. 9 is a flowchart of a training method of a lesion description model provided by an exemplary embodiment of the present application.
  • FIG. 10 is a flowchart of a method for training a benign and malignant lesion model provided by an exemplary embodiment of the present application
  • FIG. 11 is a schematic diagram of positioning of a lesion image limiting model provided by an exemplary embodiment of the present application.
  • FIG. 12 is a flowchart of a method for training a lesion matching model provided by an exemplary embodiment of the present application.
  • Fig. 13 is a working principle diagram of a breast benign and malignant model provided by an exemplary embodiment of the present application.
  • Fig. 14 is a flowchart of a method for training a single-image benign and malignant model provided by an exemplary embodiment of the present application
  • 15(A) is a working principle diagram of an automated reporting subsystem provided by another exemplary embodiment of the present application.
  • FIG. 15(B) is a schematic diagram of a structured report provided by an exemplary embodiment of the present application.
  • 16 is a schematic flow chart of a mammography target image processing method provided by an exemplary embodiment of this application.
  • FIG. 17 is a block diagram of an auxiliary detection device for mammography target images provided by an exemplary embodiment of the present application.
  • FIG. 18 is a block diagram of a computer device provided by an exemplary embodiment of the present application.
  • first, second, etc. to describe various elements, these elements should not be limited by the terms. These terms are only used to distinguish one element from another.
  • the first image may be referred to as the second image, and similarly, the second image may be referred to as the first image. Both the first image and the second image may be images, and in some cases, may be separate and different images.
  • the term “if” can be interpreted to mean “when” ("when” or “upon”) or “in response to determination” or “in response to detection.”
  • the phrase “if it is determined" or “if [the stated condition or event] is detected” can be interpreted to mean “when determining" or “in response to determining" "Or” upon detection of [statement or event]” or “in response to detection of [statement or event]”.
  • Breast molybdenum target image It is an image obtained by projecting a two-dimensional image of the breast on an X-ray film by using the physical properties of X-rays and the different iso-density values of human breast tissue, referred to as molybdenum target image.
  • the mammography target image includes: the mammography target image at the CC position and the mammography target image at the MLO position.
  • Fig. 1 shows a structural block diagram of a computer system provided by an exemplary embodiment of the present application.
  • the computer system includes: Digital Radiography (DR) equipment 02, mammography assisted diagnosis system 10 and doctor equipment 04.
  • DR Digital Radiography
  • the breast DR device 02 is a device used to collect mammography images based on digital X-ray imaging technology.
  • the mammary gland DR device 02 can belong to the mammographic mammography assisted diagnosis system 10, or it can be deployed independently of the mammary mammography assisted diagnosis system 10, as shown in FIG. 1 by taking independent deployment as an example.
  • the mammography assisted diagnosis system 10 is one or more computer devices, such as at least one of a single server, a server group, and a server cluster.
  • the system 10 includes at least one of a breast information extraction subsystem 12, a lesion identification subsystem 14, a breast benign and malignant subsystem 16, and an automated report generation subsystem 18.
  • the breast information extraction subsystem 12 is used to extract overall breast information.
  • the subsystem 12 includes: a nipple position detection model, a muscle position detection model, and a gland type classification model.
  • the nipple position detection model and the muscle position detection model can be realized as the same neural network model, that is, the nipple and muscle detection model.
  • the lesion identification subsystem 14 is used for mass detection and calcification detection on the mammography target image. The detection process will consider the type of breast glands to select the threshold.
  • the lesion recognition subsystem 14 performs a benign and malignant analysis of the detected lesions, description of lesion attributes, and matching of CC lesions and MLO lesions. In addition, the quadrant of the lesion is located according to the position of the nipple and muscle.
  • the breast benign and malignant subsystem 16 is used for predicting the benign and malignant unilateral breast based on the lesion benign and malignant analysis subsystem 14 and combining the molybdenum target image at the CC position and the molybdenum target image at the MLO position.
  • the breast benign and malignant subsystem 16 includes a breast benign and malignant detection model.
  • the breast benign and malignant detection model dynamically selects the prediction threshold according to the gland type of the breast during prediction.
  • the automated report generation subsystem 18 is used to fuse all the prediction results of the above three subsystems to generate a structured inspection report.
  • the doctor can calibrate the generated test report, and can select the region of the molybdenum target image of interest, and obtain relevant information of the molybdenum target image region through interactive query.
  • the doctor equipment 04 is a computer equipment used by a doctor, which can be a desktop computer, a notebook, a touch screen, etc.
  • the doctor device 04 is used to view the test report generated by the automated report generation subsystem 18 and the corresponding human-computer interaction.
  • the doctor equipment 04 may belong to the mammography image-assisted diagnosis system 10, or it may be deployed independently of the mammography image-assisted diagnosis system 10, as shown in FIG. 1 by taking independent deployment as an example.
  • Fig. 2 shows a flowchart of an auxiliary diagnosis method for mammography of mammography according to an exemplary embodiment of the present application.
  • This method can be applied to the auxiliary diagnosis system (hereinafter referred to as computer equipment) of mammography target images shown in FIG. 1.
  • the method includes:
  • Step 202 Obtain a mammography target image of a unilateral breast.
  • the mammography target image includes: a CC position mammography target image and an MLO position mammography target image;
  • the computer equipment obtains mammography images of a single breast in two different views.
  • the mammography target images at the two different view positions include a CC position mammography target image and an MLO position mammography target image, where the mammography target image at the CC position is a mammography target image collected in a head and tail position.
  • the center line of the CC mammography target image is from top to bottom, passing through the top of the breast and entering the center of the film vertically.
  • the image of the molybdenum target at the MLO position is an image of the molybdenum target collected at an inner oblique position.
  • the center line of the mammography target image at the MLO position is perpendicular to the center of the film through the inside of the breast.
  • the breast benign and malignant detection model is used to predict the benign and malignant mammograms of the CC position and the MLO position, and obtain the benign and malignant prediction results of the unilateral breast;
  • the breast benign and malignant detection model includes: a first single image detection unit 22, a second single image detection unit 24, a pooling layer 26, and a fully connected layer 28.
  • the computer device calls the first single image detection unit 22 to process the CC position mammography target image to obtain the first feature (such as the first logits feature); calls the second single image detection unit 24 to process the MLO position mammography target image to obtain the first feature Two features (such as the second logits feature).
  • the computer device inputs the first feature and the second feature into the pooling layer 26 and the fully connected layer 28 to obtain the benign and malignant prediction results of the unilateral breast.
  • a detection report is generated and output, the detection report includes the prediction result of benign and malignant unilateral breast.
  • the method provided in this embodiment obtains the CC mammogram image and MLO mammography target image of a unilateral breast; calls the breast benign and malignant detection model to perform good performance on the CC mammography image and MLO mammography target image.
  • Malignant prediction obtain the benign and malignant prediction results of unilateral breasts; generate and output test reports, including the benign and malignant prediction results of unilateral breasts; because the breast benign and malignant detection model can comprehensively compare the CC position mammogram and MLO position mammogram
  • the image predicts benign and malignant, that is, it combines two mammography images from different perspectives to predict benign and malignant. It more realistically simulates the doctor’s actual viewing process and improves the accuracy of predicting benign and malignant unilateral breasts. .
  • the computer device invokes the gland type classification model to identify the gland type in the mammography target image, obtains the gland type recognition result, and determines that the breast is good according to the gland type recognition result.
  • the prediction threshold corresponding to the malignant detection model is called, and the benign and malignant breast detection model after the determined prediction threshold is called to predict the benign and malignant mammogram images of the CC position and the MLO position to obtain the benign and malignant prediction results of the unilateral breast.
  • Fig. 4 shows a flowchart of an auxiliary diagnosis method for mammography of mammography provided by an exemplary embodiment of the present application. This method can be applied to the mammography assisted diagnosis system (hereinafter referred to as computer equipment) shown in FIG. 1.
  • the method includes:
  • Step 401 Obtain a mammography target image of a unilateral breast.
  • the mammography target image includes a CC position mammography target image and an MLO position mammography target image.
  • the computer equipment obtains the mammogram of the unilateral breast from the breast DR equipment. Or, the computer device receives the mammography image of the unilateral breast uploaded by the doctor device.
  • Unilateral breast refers to the left breast or the right breast.
  • the computer device obtains the mammogram of the left breast.
  • the computer device obtains the mammogram of the right breast.
  • Step 402 Invoke the gland type classification model to identify the gland type of the CC molybdenum target image, and obtain the first gland type.
  • the first gland type includes any one of fat type, small number of glands, large number of glands, and dense type.
  • Step 403 Invoke the gland type classification model to identify the gland type of the MLO molybdenum target image, and obtain the second gland type.
  • the second gland type includes any one of fat type, small number of glands, large number of glands, and dense type.
  • step 404 the first gland type and the second gland type with a higher density of glands are determined as the gland type of the unilateral breast.
  • the order of the gland density of the four gland types from largest to smallest is: dense type>multiple gland type>small number of gland type>fat type.
  • the first gland type is a gland type
  • the second gland type is a small number of glands
  • the gland type is determined as the gland type of a unilateral breast.
  • the dense type is determined as the gland type of a unilateral breast.
  • Step 405 Determine the prediction threshold corresponding to the breast benign and malignant detection model according to the gland type recognition result.
  • the computer equipment determines the corresponding prediction threshold according to the gland type recognition result.
  • the fat type corresponds to the first threshold
  • the small glandular type corresponds to the second threshold
  • the large glandular type corresponds to the third threshold
  • the dense type corresponds to the fourth threshold.
  • the computer device determines that the prediction threshold corresponding to the breast benign and malignant detection model is the first threshold.
  • the computer device determines that the prediction threshold corresponding to the breast benign and malignant detection model is the second threshold.
  • the computer device determines that the prediction threshold corresponding to the breast benign and malignant detection model is the third threshold.
  • the computer device determines that the prediction threshold corresponding to the breast benign and malignant detection model is the fourth threshold.
  • Step 406 Invoke the first single image detection unit to process the CC position mammography target image to obtain the first feature.
  • Step 407 Invoke the second single image detection unit to process the MLO position molybdenum target image to obtain the second feature.
  • Step 408 Input the first feature and the second feature into the pooling layer and the fully connected layer to obtain the benign and malignant prediction results of the unilateral breast.
  • the prediction threshold of 60% when the benign and malignant prediction probability output by the fully connected layer is 65%, since 65% is greater than 60%, it is determined that the benign and malignant prediction result is malignant.
  • the benign and malignant prediction probability output by the fully connected layer is 40%, since 40% is less than 60%, it is determined that the benign and malignant prediction result is benign.
  • the prediction threshold can be dynamically changed according to the gland type.
  • the method provided in this embodiment obtains the CC mammogram image and MLO mammography target image of a unilateral breast; calls the breast benign and malignant detection model to perform good performance on the CC mammography image and MLO mammography target image.
  • Malignant prediction obtain the benign and malignant prediction results of unilateral breasts; generate and output test reports, including the benign and malignant prediction results of unilateral breasts; because the breast benign and malignant detection model can comprehensively compare the CC position mammogram and MLO position mammogram
  • the image predicts benign and malignant, that is, it combines two mammography images from different perspectives to predict benign and malignant. It more realistically simulates the doctor’s actual viewing process and improves the accuracy of predicting benign and malignant unilateral breasts. .
  • the breast information extraction subsystem includes: nipple detection model, muscle detection model and gland type classification model.
  • the nipple detection model and the muscle detection model can be designed as the same detection model: the nipple and muscle detection model, as shown in Figure 5.
  • the nipple detection model is a two-class (nipple+background) model based on Fully Convolutional Networks (FCN).
  • FCN Fully Convolutional Networks
  • the muscle detection model is a two-class (muscle + background) model based on FCN.
  • the nipple and muscle detection model is based on the three classifications of FCN (nipple + muscle + background).
  • the computer equipment calls the nipple detection model (or the nipple and muscle detection model) to identify the position of the nipple in the mammogram.
  • the nipple detection model is used to calibrate each pixel in the mammography target image, and the type of each pixel can be either nipple or background.
  • the computer equipment calls the muscle detection model (or the nipple and muscle detection model) to identify the position of the muscle in the mammogram image.
  • the muscle detection model is used to calibrate each pixel in the mammography target image, and the type of each pixel can be either muscle or background.
  • the computer equipment calls the nipple and muscle detection model to identify the position of the nipple and the muscle in the mammogram.
  • the nipple and muscle detection model is used to calibrate each pixel in the mammography target image.
  • the type of each pixel can be one of nipple, muscle, and background.
  • the nipple and muscle detection models are used to detect the nipple area; for MLO mammography images, the nipple and muscle detection models are used to detect the nipple and muscle areas.
  • the training method of the above nipple and muscle detection model can be as follows, as shown in Figure 6:
  • Step 601 Perform weight initialization on the nipple and muscle detection model.
  • the computer device may use the FCN model issued by the University of California, Berkeley (UC Berkeley) to construct the nipple and muscle detection model, and use the segmentation data set (pattern analysis, statistical modelling and computational learning, visual object classes, PASCAL VOC) to The weights of the nipple and muscle detection models are initialized.
  • the PASCAL VOC data set is a standardized image data set for object class recognition, or a public tool set for accessing data sets and annotations.
  • Step 602 Obtain training samples.
  • the training samples include sample mammography images, nipple calibration results, and muscle calibration results.
  • the training sample includes two parts: a Digital Database for Screening Mammography (DDSM) database and a manual calibration data set.
  • the manual calibration data set can be sample mammography images (1000+) after using domestic hospital data and hiring experts to perform pixel-level standards.
  • image inversion and/or image cropping techniques can also be used for data enhancement.
  • the DDSM database is a database established by medical institutions to store breast cancer images.
  • the DDSM database stores data types such as malignant, conventional, and benign. At present, many studies on breast cancer are based on the DDSM database.
  • the first training process can use the sample mammography target images in the public data set DDSM for training, and then use the manual calibration data set for transfer learning.
  • the parameters of transfer learning can be: the input size of the sample mammography target image is 800*800 pixels, the batch size is 8, the learning rate is 0.00001, and the maximum number of iterations is 10000).
  • transfer learning is to transfer the learned and trained model parameters to a new model to help the new model training.
  • the learned model parameters can be shared with the new model in a certain way through transfer learning, so as to speed up and optimize the learning efficiency of the model instead of starting from zero like most networks.
  • Learn In this application, the network model data training is performed through parameter migration, and the model trained by task A can be used to initialize the model parameters of task B, so that task B can converge in learning and training faster.
  • Step 603 Use the nipple and muscle detection model to predict the training sample to obtain the prediction result.
  • Step 604 Perform error calculation according to the prediction result, nipple calibration result, and muscle calibration result to obtain error loss.
  • step 605 an error back propagation algorithm is used to train the nipple and muscle detection model according to the error loss.
  • the gland type classification model can be a classification model constructed by the Inception V3 model released by Google.
  • the types of glands include one of four types: fat type, small number of glands, large number of glands and dense type.
  • the Inception V3 model is a type of convolutional neural network.
  • Convolutional neural network is a feed-forward neural network. Artificial neurons can respond to surrounding units and can perform large-scale image processing.
  • Convolutional neural networks include convolutional layers and pooling layers.
  • the Inception V3 model optimizes the network by increasing the width of the single-layer convolutional layer, that is, using different scale convolution kernels on the single-layer convolutional layer.
  • the Inception V3 model approximates the optimal local sparse nodes through dense components, so that computing resources can be used more efficiently, and more features can be extracted with the same amount of calculation, thereby improving training results.
  • the convolutional layer is composed of several convolution units, and the parameters of each convolution unit are optimized through the back propagation algorithm.
  • the purpose of image convolution operation is to extract different features of the input image.
  • the first layer of convolutional layer may only extract some low-level features such as edges, lines, and corners. More layers of networks can iteratively extract from low-level features. More complex features.
  • the training method of the above-mentioned gland type classification model can be as follows, as shown in Figure 7:
  • Step 701 Perform weight initialization on the gland type classification model.
  • the computer device may use the Inception V3 model released by Google to construct a gland type classification model, and the output classification category is set to 4. Then use the ImageNet (computer vision standard data set) data set for weight initialization.
  • ImageNet computer vision standard data set
  • Step 702 Obtain training samples.
  • the training samples include sample mammography images and gland type calibration results.
  • the training sample includes two parts: the public data set DDSM released by Google Inc. and the manual calibration data set.
  • the manual calibration data set can be a sample mammography target image (1000+) after using domestic hospital data and hiring experts to calibrate the gland type.
  • image inversion and/or image cropping techniques can also be used for data enhancement.
  • the first training process can use the sample mammography target images in the public data set DDSM for training, and then use the manual calibration data set for transfer learning.
  • the parameters of migration learning can be: the error back propagation algorithm uses Root Mean Square prop (RMSprop), the batch size is 64, the initial learning rate is 0.00001, and the maximum number of iterations is 10,000.
  • RMSprop Root Mean Square prop
  • Step 703 Use the gland type classification model to predict the training sample to obtain the prediction result.
  • Step 704 Perform error calculation according to the prediction result and the gland type calibration result to obtain the error loss.
  • step 705 an error back propagation algorithm is used to train the gland type classification model according to the error loss.
  • the lesion recognition subsystem includes: lesion description model, lesion benign and malignant model, lesion matching model and lesion image limiting model, as shown in Figure 8.
  • the computer equipment calls the lesion description model to detect the lesions in the mammography target image, and obtains lesion description information.
  • the computer equipment invokes the benign and malignant lesion model to identify the benign and malignant lesions in the mammography target image, and obtain the benign and malignant probability of the lesion.
  • the computer equipment calls the lesion matching model to judge the consistency of the lesions of the CC mammography target image and the MLO mammography target image, and obtain the lesion matching probability.
  • the computer equipment invokes the limited location model of the lesion image to calculate the quadrant of the lesion in the mammography target image.
  • the lesion description model can be a classification model constructed by the Inception V3 model released by Google.
  • the training method of the above lesion description model can be as follows, as shown in Figure 9:
  • Step 901 Perform weight initialization on the lesion description model.
  • the Inception V3 model released by Google is used to construct the lesion description model.
  • the last fully connected layer of the Inception V3 model is modified to support multiple parallel fully connected layers for simultaneous training of multiple tasks, and the output category corresponding to each task is set to 2, that is, shared by each task All parameters except the last fully connected layer. Then use the ImageNet dataset for weight initialization.
  • Step 902 Obtain training samples.
  • the training samples include sample mammography images and lesion calibration results.
  • the training sample includes two parts: the public data set DDSM released by Google Inc. and the manual calibration data set.
  • the manual calibration data set can be a sample mammography target image (1000+) after using domestic hospital data and hiring experts to calibrate the attributes of the lesion. Take the lump attribute as an example. Mark each lump with a round or irregular shape, a clear or fuzzy boundary, a paging or no paging on the boundary, and a burr or no burr on the boundary.
  • image inversion and/or image cropping techniques can also be used for data enhancement.
  • the first training process can use the sample mammography target images in the public data set DDSM for training, and then use the manual calibration data set for transfer learning.
  • the parameters of migration learning can be: the error back propagation algorithm uses Adam, the batch size is 64, the initial learning rate is 0.001, and the maximum number of iterations is 10,000.
  • Step 903 Use the lesion description model to predict the training sample to obtain the prediction result.
  • Step 904 Perform error calculation according to the prediction result and the lesion calibration result to obtain the error loss.
  • step 905 an error back propagation algorithm is used to train the lesion description model according to the error loss.
  • the benign and malignant lesion model can be a classification model constructed by the Inception V3 model released by Google.
  • the training method of the above-mentioned benign and malignant lesion model can be as follows, as shown in Figure 10:
  • Step 1001 Perform weight initialization on the benign and malignant lesion model.
  • the Inception V3 model released by Google is used to construct a benign and malignant lesion model.
  • the last pooling layer of the Inception V3 model is modified to a maximum pooling layer (max pooling), and the number of output categories is set to 2. Then use the ImageNet dataset for weight initialization.
  • Step 1002 Obtain training samples.
  • the training samples include sample mammography images and benign and malignant calibration results.
  • the training sample includes two parts: the public data set DDSM released by Google Inc. and the manual calibration data set.
  • the manual calibration data set can be a sample mammography target image (16000+) after using domestic hospital data and hiring experts to calibrate the benign and malignant lesions.
  • the benign and malignant calibration results include: malignant calcification lesions and malignant mass lesions are positive samples, and benign calcification lesions, benign mass lesions and normal areas are negative samples.
  • image inversion and/or image cropping techniques can also be used for data enhancement.
  • the first training process can use the sample mammography target images in the public data set DDSM for training, and then use the manual calibration data set for transfer learning.
  • the parameters of migration learning can be: the error back propagation algorithm uses Adam, the batch size is 64, the initial learning rate is 0.001, and the maximum number of iterations is 10,000.
  • Step 1003 Predict the training samples using the benign and malignant lesion model to obtain the prediction result.
  • Step 1004 Perform error calculation according to the prediction result and the lesion calibration result to obtain the error loss.
  • Step 1005 According to the error loss, the error back propagation algorithm is used to train the benign and malignant lesion model.
  • the training end condition is satisfied.
  • the lesion image limiting model is a model based on the muscle line fitting equation.
  • the muscle boundary line equation is obtained by linear fitting the boundary of the muscle position in the mammogram image. Then, according to the position of the nipple, quadrant calculation of the lesion in the mammography target image is performed.
  • the computer equipment obtains the limited position model of the lesion image.
  • the limited position model of the lesion image is obtained through training of multiple sample images. Each pixel in the sample image is marked.
  • the annotation types include: background, nipple, and muscle.
  • the limited location model of the lesion image can identify that each pixel in an image belongs to either the background, the nipple or the muscle.
  • the computer equipment inputs the CC mammogram image into the lesion image limitation model, and based on the lesion image limitation model, the nipple position in the CC mammography image can be determined (the CC mammography image has no muscle information, so there is no muscle area); Input the MLO mammography target image into the lesion image limitation model, and based on the lesion image limitation model, the nipple position and muscle position in the MLO mammography image can be determined.
  • the computer device determines the first segmentation line according to the position of the nipple and the boundary line of the breast border, and determines that the first lesion area is located in the inner quadrant or the outer quadrant according to the first segmentation line.
  • the computer equipment fits the muscle boundary line equation according to the muscle position, and then determines the muscle boundary line (that is, the object boundary line in the foregoing), and then determines the second according to the nipple position and muscle boundary line
  • the dividing line according to the second dividing line, determines that the second lesion area is located in the upper quadrant or the lower quadrant.
  • the computer device 10a acquires the mammography image of the same breast of the same patient, and displays the acquired mammography image on the screen, where the mammography image includes :CC position mammography target image 20b, and MLO position mammography target image 20c.
  • the CC position mammography target image is to image the breast according to the head and tail position
  • the MLO position mammography image is to image the breast according to the oblique lateral position.
  • the computer device 10a obtains a mass detection model and a calcification detection model (that is, the lesion description model in the present application).
  • the mass detection model can identify the location area of the tumor in the image; the calcification detection model can identify the location of the calcification focus in the image The location area, mass category and calcification category belong to the lesion category.
  • the computer device 10a can input the CC mammography image 20b into the mass detection model, and the mass detection model can output the CC mammography image 20b.
  • the lesion object in the CC mammography image 20b is located in the lesion area 20d in the CC mammography image 20b.
  • the lesion category to which the lesion object in the CC mammography image 20b belongs is a mass category.
  • the computer device 10a can input the MLO mammogram image 20c into the mass detection model, and the mass detection model can also output the MLO mammogram image 20c.
  • the lesion object in the MLO mammogram image 20c is located in the lesion area 20h in the MLO mammogram image 20c, and the MLO can also be determined.
  • the lesion category to which the lesion object in the mammography image 20c belongs is a mass category.
  • the computer device 10a also inputs the CC molybdenum target image 20b into the calcification detection model, and the calcification detection model does not detect calcification in the CC molybdenum target image 20b; the computer device 10a also inputs the MLO molybdenum target image Enter the calcification detection model at 20c.
  • the calcification detection model also did not detect calcification lesions in the mammogram 20c at the MLO position.
  • the tumor is located in the lesion area 20d; in the MLO mammography image 20c, the mass is The lesion was located in the lesion area for 20h.
  • the computer device 10a obtains a limited position model of the lesion image.
  • the limited position model of the lesion image may identify the tissue category to which each pixel in the image belongs.
  • the tissue categories include: nipple category, muscle category, and background category.
  • the computer device 10a inputs the CC mammogram image 20b into the lesion image limitation model, and the model can determine the tissue category to which each pixel of the CC mammography image 20b belongs.
  • the computer device 10a combines the pixels belonging to the nipple category into an area 20e, which is the area where the nipple is located.
  • the computer device 10a determines the breast edge line 20g in the CC position mammography image 20b, and uses the line 20f perpendicular to the breast edge line 20g and passing through the region 20e as the quadrant division line 20f.
  • those located below the quadrant dividing line 20f are the inner quadrants, and those located above the quadrant dividing line 20f are the outer quadrants.
  • the computer device 10a can determine the CC mammography image
  • the mass lesion in 20b is located in the inner quadrant.
  • the computer device 10a inputs the MLO mammogram image 20c into the lesion image limitation model, and the model can determine the tissue category to which each pixel of the MLO mammography image 20c belongs.
  • the computer device 10a combines the pixels belonging to the nipple category into an area 20j, which is the area where the nipple is located.
  • the computer device 10a combines the quadrants belonging to the muscle category into a muscle area, determines the area boundary line 20m between the muscle area and the non-muscle area, and uses a line 20k perpendicular to the area boundary line 20m and passing through the area 20j as the quadrant division line 20k.
  • the lower quadrant is located below the quadrant division line 20k, and the upper quadrant is located above the quadrant division line 20k.
  • the computer device 10a can determine that the tumor in the MLO mammography image 20c is located in the lower quadrant.
  • the computer device 10a combines the inner quadrant determined by the CC position mammography target image 20b and the lower quadrant determined by the MLO position mammography target image 20c into quadrant position information 20n "inner lower quadrant".
  • the computer device 10a can combine the quadrant position information 20n "inner lower quadrant” and the lesion type "lumps" corresponding to the CC mammography image 20b and MLO mammography image 20c into a diagnosis opinion: "lumps are seen in the inner lower quadrant” .
  • the lesion matching model may be a classification model constructed based on the VGG model released by the University of Oxford.
  • the training process of the lesion matching model can be as follows, as shown in Figure 12:
  • Step 1201 Perform weight initialization on the lesion matching model.
  • the VGG model released by Oxford University is used to construct the lesion matching model.
  • the pool5 layer results of the VGG model are taken for fusion, and then three fully connected layers are used in accordance with the original VGG to obtain a classification number of 2 results. Then use the ImageNet dataset for weight initialization.
  • Step 1202 Obtain training samples.
  • the training samples include sample mammography images and matching calibration results.
  • the training sample includes two parts: the public data set DDSM released by Google Inc. and the manual calibration data set.
  • the manual calibration data set can be the use of domestic hospital data, and hiring experts to examine a pair of plaques representing the same lesion in the CC mammography image and MLO mammography image as positive samples, and any other two that do not represent the same lesion. Take the plaque as a negative sample, and get the sample mammogram (1000+).
  • image inversion and/or image cropping techniques can also be used for data enhancement.
  • the first training process can use the sample mammography target images in the public data set DDSM for training, and then use the manual calibration data set for transfer learning.
  • the parameters of migration learning can be: the error back propagation algorithm uses Adam, the batch size is 128, the initial learning rate is 0.001, and the maximum number of iterations is 10,000.
  • Step 1203 Use the lesion matching model to predict the training sample to obtain the prediction result.
  • Step 1204 Perform error calculation according to the prediction result and the matching calibration result to obtain the error loss.
  • Step 1205 According to the error loss, an error back propagation algorithm is used to train the lesion matching model.
  • the attribute classification network used to identify the degree of lesion matching is obtained by training.
  • the breast benign and malignant subsystem includes: a breast benign and malignant model, as shown in Figure 13.
  • the breast benign and malignant model is used to detect the benign and malignant images of the CC and MLO mammograms of a unilateral breast to obtain the benign and malignant probability of a unilateral breast.
  • the structure of the breast benign and malignant model is shown in Figure 3.
  • the training process of the above-mentioned benign and malignant breast model includes: training of a single image benign and malignant model, and training of a benign and malignant breast model.
  • the single-image benign and malignant model is used to construct the first single-image detection unit 22 and the second single-image detection unit 24 in FIG. 3.
  • the training methods of single-image benign and malignant models include the following:
  • Step 1401 Obtain the trained benign and malignant lesion model as an initial single-image benign and malignant model.
  • Step 1402 Obtain training samples.
  • the training samples include the sample mammography target image and the whole image benign and malignant calibration results;
  • the training sample includes two parts: a manual calibration data set.
  • the manual calibration data set can be a sample mammogram (16000+) after using domestic hospital data and hiring experts to calibrate the benign and malignant aspects of the entire picture.
  • the benign and malignant calibration results of the whole image include: a malignant mammogram image as a positive sample, and the whole image is a benign and/or normal mammogram image as a negative sample.
  • image inversion and/or image cropping techniques can also be used for data enhancement.
  • the parameters of migration learning can be: the error back propagation algorithm uses Adam, the batch size is 64, the initial learning rate is 0.001, and the maximum number of iterations is 10,000.
  • Step 1403 Use the single-image benign and malignant model to predict the training sample to obtain the prediction result.
  • Step 1404 Perform error calculation according to the prediction result and the benign and malignant calibration result of the whole image to obtain the error loss.
  • step 1405 an error back propagation algorithm is used to train the single-image benign and malignant model according to the error loss.
  • the training end condition is satisfied. Trained to obtain an attribute classification network for identifying benign and malignant single images. In some embodiments, a probability greater than 0.5 is considered to be suspected of containing a malignant lesion.
  • the pooling layer 25 and the fully connected layer 28 are added to keep the output category of the model at 2. , Forming a benign and malignant breast model.
  • the training samples of the single-image benign and malignant model are used as the new training samples.
  • After data enhancement because it is a molybdenum target image, the data is mainly enhanced by flipping and cropping, there is no need to perform data enhancement in color space).
  • the malignant classification model is extended to the CC and MLO dual-map benign and malignant classification model.
  • the training parameters can be: (the descent algorithm uses RMSprop, the batch size is 32, the initial learning rate is 0.01, and the maximum number of iterations is 10000).
  • Figure 15(A) shows a schematic diagram of the principle of the automatic report generation subsystem fusing the results of the above-mentioned subsystems and automatically generating a structured report.
  • the automatic report generation subsystem generates a structured report by fusing all the detection and recognition results of the above-mentioned subsystems.
  • the doctor can modify the generated report and query relevant information about the area of interest in an interactive manner. The detailed description of each step is as follows:
  • the automatic report generation subsystem has the following functions:
  • the computer equipment summarizes the detection and identification results obtained by all the above subsystems, and automatically generates the report content of the detection report described in the BI-RADS standard.
  • the report content of the test report includes: description of mass, description of calcification, description of gland type, description of breast benign and malignant, etc.
  • the test report can refer to FIG. 15(B).
  • the test report includes: a small number of glandular types, malignant masses, malignant calcifications, high probability of malignancy in the right breast, masses in the inner and lower quadrants, and description information of the masses.
  • the description information of the masses includes: irregular shapes, unclear borders, and At least one of glitches and shallow paging.
  • Doctors can revise the system's detection and recognition results and automatically generated reports through review and other methods to obtain a diagnosis report.
  • the computer device receives the report correction request sent by the doctor device, and corrects the test report according to the report correction request.
  • Doctors can query the information they are interested in interactively. For example, the doctor can select the region of interest in the mammogram image, and obtain the benign and malignant results of the region by calling the benign and malignant lesion classification model of the system.
  • the computer device also receives a local query request, which is used to request a query for a local area in the mammography target image, and output a detection report corresponding to the local area according to the local query request.
  • the present application also provides a mammography target image processing method, which is executed by a computer device, and the mammography target image processing method includes the following steps:
  • Step 1602 Obtain a mammogram image of a unilateral breast.
  • the mammogram image includes: a head and tail CC image and a medial oblique MLO image.
  • Step 1604 Invoke the target detection model to process the CC mammography target image and the MLO mammography target image to obtain an image detection result of the unilateral breast.
  • Step 1606 Generate and output a detection report according to the image detection result.
  • the target detection model includes: a first single-image detection part, a second single-image detection part, a pooling layer, and a fully connected layer.
  • the step of calling the target detection model to process the CC molybdenum target image and the MLO molybdenum target image to obtain the image detection result of the unilateral breast specifically includes: invoking the first single image detection unit to the CC molybdenum target image Perform processing to obtain the first feature; call the second single image detection unit to process the MLO target image to obtain the second feature; and input the first feature and the second feature into the pooling layer and the global Connect the layers to obtain the image detection result of the unilateral breast.
  • the mammography image processing method further includes: calling gland The body type classification model recognizes the gland type in the mammography target image, and obtains the gland type recognition result.
  • the step of calling the target detection model to process the CC molybdenum target image and the MLO molybdenum target image to obtain the image detection result of the unilateral breast specifically includes: determining the prediction corresponding to the target detection model according to the gland type recognition result Threshold; and calling the target detection model after determining the predicted threshold to process the CC molybdenum target image and the MLO molybdenum target image to obtain an image detection result of the unilateral breast.
  • calling the gland type classification model to identify the gland type in the mammography target image and obtaining the gland type recognition result specifically includes: calling the gland type classification model for the CC molybdenum target Identify the gland type of the image to obtain the first gland type; call the gland type classification model to identify the gland type of the MLO mammogram image to obtain the second gland type; and the first gland type The type and the second gland type with a higher density of glands are determined as the gland type of the unilateral breast.
  • the mammography target image processing method further includes: calling an abnormality recognition model to perform abnormality detection on the mammography target image to obtain an abnormality detection result.
  • the abnormality detection includes at least one of mass detection and calcification detection.
  • the anomaly recognition model includes at least one of an anomaly description model, an anomaly classification model, an anomaly matching model, and an abnormal image limiting model; the anomaly identification model is called to perform anomaly detection on the mammography target image to obtain
  • the step of abnormal detection results specifically includes at least one of the following steps: calling the abnormal description model to detect the abnormal area in the mammography target image to obtain abnormal description information; calling the abnormal classification model to the mammography target image Recognize the abnormal category in the database to obtain the corresponding category probability; call the abnormal matching model to determine the consistency of the abnormal area of the CC mammography target image and the MLO mammography target image to obtain the abnormal area matching probability; and call the abnormality
  • the image limiting model performs quadrant calculation on the abnormal area in the mammography target image.
  • the mammography image processing method before calling the abnormal image limiting position model to perform quadrant calculation on the abnormal area in the mammography target image, the mammography image processing method further includes: calling the nipple detection model for the nipple in the mammography target image. Location for identification.
  • the step of calling the abnormal image limiting position model to perform quadrant calculation of the abnormal region in the mammography target image specifically includes: calling the abnormal image limiting position model to perform quadrant calculation of the abnormal region in the mammography target image according to the nipple position.
  • the mammography image processing method further includes: when the mammography image is an MLO mammography image, calling a muscle detection model to identify the muscle position in the mammography image.
  • the mammography image processing method further includes: receiving a report correction request; and correcting the detection report according to the report correction request.
  • the mammography image processing method further includes: receiving a local query request, the local query request is used to request a query for a local area in the mammography image; and outputting the local area corresponding to the local query request Test report.
  • the specific processing process of the target detection model can refer to the processing process of the breast benign and malignant detection model in the auxiliary diagnosis method of breast molybdenum target image mentioned in the foregoing embodiment
  • the specific processing process of the abnormal recognition model can refer to the foregoing embodiment
  • the specific processing process of the abnormal description model can refer to the description of the lesion description model in the method for assisted diagnosis of mammography target images mentioned in the previous embodiment.
  • the specific processing process of the abnormal classification model can refer to the processing process of the benign and malignant lesion model in the auxiliary diagnosis method of mammography target images mentioned in the previous embodiment
  • the specific processing process of the abnormal matching model can refer to the previous embodiment
  • the specific processing process of the abnormal image limitation model can refer to the lesion image limitation in the assisted diagnosis method of breast mammography image mentioned in the previous embodiment
  • Fig. 17 shows a block diagram of an auxiliary diagnosis device for mammography images provided by an exemplary embodiment of the present application.
  • the device can be used to realize the function of the auxiliary diagnosis method of the mammography target image of the breast.
  • the device includes:
  • the image acquisition module 101 is configured to acquire a mammography target image of a unilateral breast.
  • the mammography target image includes a CC position mammography target image and an MLO position mammography target image.
  • the breast benign and malignant detection model 102 is used for predicting the benign and malignant mammograms of the CC position and the MLO position of mammography to obtain the benign and malignant prediction results of the unilateral breast.
  • the automatic report generation module 103 is used to generate and output a detection report, the detection report including the benign and malignant prediction results of the unilateral breast.
  • the breast benign and malignant detection model 102 includes: a first single image detection part, a second single image detection part, a pooling layer, and a fully connected layer; the breast benign and malignant detection model 102 is used to call the A single image detection unit processes the CC-bit molybdenum target image to obtain the first logits feature.
  • the breast benign and malignant detection model 102 is used to call the second single-image detection unit to process the MLO mammogram image to obtain the second logits feature.
  • the breast benign and malignant detection model 102 is used to input the first logits feature and the second logits feature into the pooling layer and the fully connected layer to obtain a benign and malignant prediction result of the unilateral breast.
  • the auxiliary diagnosis device for mammography mammography images further includes: a gland type classification model.
  • the gland type classification model is used to identify the gland type in the mammography target image to obtain the gland type recognition result.
  • the breast benign and malignant detection model 102 is used to determine the prediction threshold corresponding to the breast benign and malignant detection model according to the recognition result of the gland type; call the breast benign and malignant detection model after determining the prediction threshold to the CC molybdenum target image and the The benign and malignant prediction results of the unilateral breast are obtained by using the MLO mammography target image.
  • the gland type classification model is used to call the gland type classification model to identify the gland type of the CC molybdenum target image to obtain the first gland type; call the gland type classification model Identify the gland type of the mammography target image at the MLO position to obtain the second gland type; determine the one with the greater density of the first gland type and the second gland type as the unilateral The type of glands in the breast.
  • the auxiliary diagnosis device for the mammography target image further includes: a lesion recognition model; the lesion recognition model is used to perform lesion detection on the mammography target image to obtain a lesion detection result, and the lesion detection includes mass detection And at least one of calcification detection.
  • the lesion recognition model includes at least one of a lesion description model, a benign and malignant lesion model, a lesion matching model, and a lesion image limiting model;
  • the lesion description model is used to detect the lesion in the mammography target image of the breast to obtain lesion description information.
  • the benign and malignant lesion model is used to identify benign and malignant lesions in the mammogram of the breast to obtain the benign and malignant probability of the lesion.
  • the lesion matching model is used to judge the consistency of the lesions of the CC mammography target image and the MLO mammography target image to obtain the lesion matching probability.
  • the lesion image limiting model is used to calculate the quadrant of the lesion in the mammography target image.
  • the auxiliary diagnosis device for the mammogram image further includes: a nipple detection model; the nipple detection model is used to identify the position of the nipple in the mammogram image.
  • the lesion image limiting model is used to calculate the quadrant of the lesion in the mammogram of the breast according to the position of the nipple.
  • the auxiliary diagnosis device for the mammography image of the breast further includes: a muscle detection model; the muscle detection model is used for when the mammography image of the mammogram is an MLO position mammography image, To identify the position of the muscle.
  • the automated report generation module 103 is configured to receive a report correction request; and correct the detection report according to the report correction request.
  • the automated report generation module 103 is configured to receive a local query request, the local query request is used to request a query of a local area in the mammography image; according to the local query request, the detection corresponding to the local area is output report.
  • the application also provides a mammography mammography target image processing device, which can be used to realize the functions of the mammography mammography image processing method described above.
  • the mammography target image processing device includes:
  • the image acquisition module is used to acquire mammography mammography images of one breast.
  • the mammography mammography images include CC mammography images and MLO mammography images.
  • the target detection model is used to process the CC molybdenum target image and the MLO molybdenum target image to obtain the image detection result of the unilateral breast.
  • the automatic report output module is used to generate and output a test report according to the image test result.
  • the target detection model includes: a first single-image detection part, a second single-image detection part, a pooling layer, and a fully connected layer.
  • the model detection model is used to call the first single image detection unit to process the CC molybdenum target image to obtain the first logits feature.
  • the target detection model is used to call the second single-image detection unit to process the MLO molybdenum target image to obtain the second logits feature.
  • the breast benign and malignant detection model is used to input the first logits feature and the second logits feature into the pooling layer and the fully connected layer to obtain an image detection result of a unilateral breast.
  • the mammogram image processing device further includes: a gland type classification model.
  • the gland type classification model is used to identify the gland type in the mammography target image to obtain the gland type recognition result.
  • the target detection model is used to determine the prediction threshold corresponding to the target detection model according to the recognition result of the gland type; call the target detection model after determining the prediction threshold to process the CC molybdenum target image and the MLO molybdenum target image , Get the image detection result of the unilateral breast.
  • the gland type classification model is used to call the gland type classification model to identify the gland type of the CC molybdenum target image to obtain the first gland type; call the gland type classification model Identify the gland type of the mammography target image at the MLO position to obtain the second gland type; determine the one with the greater density of the first gland type and the second gland type as the unilateral The type of glands in the breast.
  • the mammography image processing device further includes an abnormal recognition model.
  • the abnormal recognition model is used to perform abnormal detection on the mammography target image to obtain abnormal detection results, and the abnormal detection includes at least one of mass detection and calcification detection.
  • the anomaly recognition model includes at least one of an anomaly description model, an anomaly classification model, an anomaly matching model, and an anomaly identification model.
  • the abnormal description model is used to detect abnormal regions in the mammography target image to obtain abnormal description information.
  • the abnormal classification model is used to identify the abnormal category in the mammography target image to obtain the category probability.
  • the abnormal matching model is used for judging the consistency of the abnormal region of the CC mammography target image and the MLO mammography target image to obtain the matching probability of the abnormal region.
  • This abnormal image limiting model is used to calculate the quadrant of the abnormal area in the mammography target image.
  • the mammography image processing device further includes: a nipple detection model.
  • the nipple detection model is used to identify the position of the nipple in the mammography target image.
  • the abnormal image limiting position model is used for quadrant calculation of the abnormal region in the mammography target image according to the position of the nipple.
  • the mammography image processing device further includes a muscle detection model.
  • the muscle detection model is used to identify the position of the muscle in the mammography target image when the mammography target image is an MLO position mammography target image.
  • the automated report generation module 103 is configured to receive a report correction request; and correct the detection report according to the report correction request.
  • the automated report generation module 103 is configured to receive a local query request, the local query request is used to request a query of a local area in the mammography image; according to the local query request, the detection corresponding to the local area is output report.
  • Fig. 18 shows a schematic structural diagram of a computer device provided by an exemplary embodiment of the present application.
  • the computer device 1800 includes a central processing unit (Central Processing Unit, CPU for short) 1801, including random access memory (random access memory, RAM for short) 1802 and read-only memory (read-only memory, for short). : ROM) 1803, the system memory 1804, and the system bus 1805 connecting the system memory 1804 and the central processing unit 1801.
  • the computer equipment 1800 also includes a basic input/output system (I/O system) 1806 that helps to transfer information between various devices in the computer, and a large-capacity storage system 1813, a client 1814, and other program modules 1815.
  • the basic input/output system 1806 includes a display 1808 for displaying information and an input device 1809 such as a mouse and a keyboard for the user to input information.
  • the display 1808 and the input device 1809 are both connected to the central processing unit 1801 through the input/output controller 1180 connected to the system bus 1805.
  • the basic input/output system 1806 may also include an input/output controller 1180 for receiving and processing input from multiple other devices such as a keyboard, a mouse, or an electronic stylus.
  • the input/output controller 1180 also provides output to a display screen, a printer, or other types of output devices.
  • the mass storage device 1807 is connected to the central processing unit 1801 through a mass storage controller (not shown) connected to the system bus 1805.
  • the mass storage device 1807 and its associated computer-readable medium provide non-volatile storage for the computer device 1800. That is, the mass storage device 1807 may include a computer readable medium (not shown) such as a hard disk or a compact disc read-only memory (Compact Disc Read-Only Memory, CD-ROM for short) drive.
  • a computer readable medium such as a hard disk or a compact disc read-only memory (Compact Disc Read-Only Memory, CD-ROM for short) drive.
  • Computer-readable media may include computer storage media and communication media.
  • Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storing information such as computer readable instructions, data structures, program modules or other data.
  • Computer storage media include RAM, ROM, erasable programmable read-only memory (erasable programmable read-only memory, abbreviated as EPROM), electrically erasable programmable read-only memory (electrically erasable programmable read-only memory, abbreviated as EEPROM) ), flash memory or other solid-state storage technology, CD-ROM, Digital Versatile Disc (Digital Versatile Disc, DVD) or other optical storage, tape cartridges, magnetic tape, magnetic disk storage or other magnetic storage devices.
  • EPROM erasable programmable read-only memory
  • EEPROM electrically erasable programmable read-only memory
  • flash memory or other solid-state storage technology CD-ROM, Digital Versatile Disc (Digital Versatile Disc, DVD) or other optical storage, tape cartridges, magnetic tape, magnetic disk storage or other magnetic storage devices.
  • CD-ROM Compact Disc
  • DVD Digital Versatile Disc
  • tape cartridges magnetic tape
  • magnetic disk storage magnetic disk storage devices
  • the computer device 1800 may also be connected to a remote computer on the network through a network such as the Internet to run. That is, the computer device 1800 can be connected to the network 1812 through the network interface unit 1811 connected to the system bus 1805, or in other words, the network interface unit 1811 can also be used to connect to other types of networks or remote computer systems (not shown) ).
  • a computer device including a memory and a processor, where computer-readable instructions are stored in the memory, and the processor implements the steps in the foregoing method embodiments when executing the computer-readable instructions.
  • a computer-readable storage medium which stores computer-readable instructions, and the computer-readable instructions implement the steps in the foregoing method embodiments when executed by a processor.
  • an auxiliary diagnosis system for mammography of mammography includes: breast DR equipment, computer equipment, and doctor equipment; the breast DR equipment is connected to the computer equipment, and the computer equipment is connected to the doctor equipment.
  • the present application also provides a computer program product containing instructions, which when run on a computer device, causes the computer device to execute the mammography assisted diagnosis method provided by each of the foregoing method embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Public Health (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Pathology (AREA)
  • Quality & Reliability (AREA)
  • Databases & Information Systems (AREA)
  • Surgery (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Optics & Photonics (AREA)
  • Animal Behavior & Ethology (AREA)
  • Veterinary Medicine (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Dentistry (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

一种乳腺钼靶图像处理方法，由计算机设备执行，所述方法包括：获取单侧乳房的乳腺钼靶图像，所述乳腺钼靶图像包括：CC位钼靶图像和MLO位钼靶图像；调用目标检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果；及根据所述图像检测结果生成检测报告并输出。

Description

乳腺钼靶图像处理方法、装置、系统及介质
本申请要求于2019年05月16日提交中国专利局，申请号为201910407807.X、发明名称为“乳腺钼靶图像的辅助诊断方法、装置、系统及介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及人工智能领域，特别涉及一种乳腺钼靶图像处理方法、装置、系统及介质。
背景技术
乳腺钼靶(mammograms)图像被广泛应用于乳腺癌早期筛查，医生通常会诊断钼靶图像中的各种异常信息，包括肿块病灶、钙化病灶、淋巴结肿大及皮肤异常等，并根据异常信息给出表征风险程度的乳腺影像报告和数据系统(BI-RADS)分级。
相关技术中，采用神经网络模型对乳腺钼靶图像中的异常区域进行识别，以定位出乳腺钼靶图像中的异常区域并获取相应的异常信息，从而根据乳腺钼靶图像中的所有异常信息推断出该乳腺钼靶图像整体的BI-RADS评分。
但是上述神经网络模型的预测过程仅是简单的二维图像层面的预测过程，与医生实际的看片过程并不吻合，导致对异常区域的识别准确程度较低。
发明内容
本申请的各种实施例提供了一种乳腺钼靶图像的辅助诊断方法、装置、计算机设备、系统及介质，一种乳腺钼靶图像处理方法、装置、计算机设备、系统及介质。
一种乳腺钼靶图像的辅助诊断方法，由计算机设备执行，所述方法包括：
获取单侧乳房的乳腺钼靶图像，所述乳腺钼靶图像包括：头尾(Cranial-Caudal，CC)位钼靶图像和内侧斜视(MedioLateral-Oblique，MLO)位钼靶图像；
调用乳房良恶性检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行良恶性预测，得到所述单侧乳房的良恶性预测结果；及
生成和输出检测报告，所述检测报告包括所述单侧乳房的良恶性预测结果。
一种乳腺钼靶图像的辅助诊断装置，所述装置包括：
图像获取模块，用于获取单侧乳房的乳腺钼靶图像，所述乳腺钼靶图像包括：CC位钼靶图像和MLO位钼靶图像；
乳房良恶性检测模型，用于对所述CC位钼靶图像和所述MLO位钼靶图像进行良恶性预测，得到所述单侧乳房的良恶性预测结果；及
自动化报告输出模块，用于生成和输出检测报告，所述检测报告包括所述单侧乳房的良恶性预测结果。
一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现如上所述的乳腺钼靶图像的辅助诊断方法的步骤。
一种乳腺钼靶图像的辅助诊断系统，所述系统包括：乳腺DR设备、计算机设备和医生设备；所述乳腺DR设备与所述计算机设备相连，所述计算机设备与所述医生设备相连；所述计算机设备包括存储器和处理器；所述存储器存储有计算机可读指令，所述计算机可读指令由所述处理器加载并执行以实现如上所述的乳腺钼靶图像的辅助诊断方法。
一种计算机可读存储介质，所述存储介质中存储有计算机可读指令，所述计算机可读指令由处理器加载并执行以实现如上所述的乳腺钼靶图像的辅助诊断方法的步骤。
一种乳腺钼靶图像处理方法，由计算机设备执行，所述方法包括：
获取单侧乳房的乳腺钼靶图像，所述钼靶图像包括：头尾CC位钼靶图 像和内侧斜MLO位钼靶图像；
调用目标检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果；及
根据所述图像检测结果生成检测报告并输出。
一种乳腺钼靶图像处理装置，所述装置包括：
图像获取模块，用于获取单侧乳房的乳腺钼靶图像，所述乳腺钼靶图像包括：头尾CC位钼靶图像和内侧斜MLO位钼靶图像；
目标检测模型，用于对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果；及
自动化报告输出模块，用于根据所述图像检测结果生成检测报告并输出。
一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现所述乳腺钼靶图像处理方法的步骤。
一种乳腺钼靶图像处理系统，所述系统包括：乳腺DR设备、计算机设备和医生设备；所述乳腺DR设备与所述计算机设备相连，所述计算机设备与所述医生设备相连；所述计算机设备包括存储器和处理器，所述存储器存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现所述乳腺钼靶图像处理方法的步骤。
一种计算机可读存储介质，存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现如上所述的乳腺钼靶图像处理方法的步骤。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其它特征、目的和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本 申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本申请一个示意性实施例提供的乳腺钼靶图像的辅助检测方法的流程图；
图2是本申请另一个示意性实施例提供的乳腺钼靶图像的辅助检测方法的流程图；
图3是本申请一个示意性实施例提供的乳房良恶性模型的结构框图；
图4是本申请另一个示意性实施例提供的乳腺钼靶图像的辅助检测方法的流程图；
图5是本申请另一个示意性实施例提供的乳房信息提取子系统的流程图；
图6是本申请另一个示意性实施例提供的乳头和肌肉检测模型的训练方法的流程图；
图7是本申请另一个示意性实施例提供的腺体类型分类模型的训练方法的流程图；
图8是本申请另一个示意性实施例提供的病灶检测子系统的工作原理图；
图9是本申请一个示意性实施例提供的病灶描述模型的训练方法的流程图；
图10是本申请一个示意性实施例提供的病灶良恶性模型的训练方法的流程图；
图11是本申请一个示意性实施例提供的病灶象限定位模型的定位示意图；
图12是本申请一个示意性实施例提供的病灶匹配模型的训练方法的流程图；
图13是本申请一个示意性实施例提供的乳房良恶性模型的工作原理图；
图14是本申请一个示意性实施例提供的单图良恶性模型的训练方法的 流程图；
图15(A)是本申请另一个示意性实施例提供的自动化报告子系统的工作原理图；
图15(B)是本申请一个示意性实施例提供的结构化报告的报告示意图；
图16为本申请一个示意性实施例所提供的乳腺钼靶图像处理方法的流程示意图；
图17是本申请一个示意性实施例提供的乳腺钼靶图像的辅助检测装置的框图；
图18是本申请一个示意性实施例提供的计算机设备的框图。
具体实施方式
为使本申请实施例的目的、技术方案和优点更加清楚，下面将结合附图对本申请实施方式作进一步地详细描述。
尽管以下描述使用术语第一、第二等来描述各种元素，但这些元素不应受术语的限制。这些术语只是用于将一元素与另一元素区域分开。例如，在不脱离各种所述示例的范围的情况下，第一图像可以被称为第二图像，并且类似地，第二图像可以被称为第一图像。第一图像和第二图像都可以是图像，并且在某些情况下，可以是单独且不同的图像。
在本文中对各种所述示例的描述中所使用的术语只是为了描述特定示例，而并非旨在进行限制。如在对各种所述示例的描述和所附权利要求书中所使用的那样，单数形式“一个(“a”“,an”)”和“该”旨在也包括复数形式，除非上下文另外明确地指示。还将理解的是，本文中所使用的术语“和/或”是指并且涵盖相关联的所列出的项目中的一个或多个项目的任何和全部可能的组合。还将理解的是，术语“包括”(“inCludes”“inCluding”“Comprises”和/或“Comprising”)当在本说明书中使用时指定存在所陈述的特征、整数、步骤、操作、元素、和/或部件，但是并不排除存在或添加一个或多个其他特征、整数、步骤、操作、元素、部件、和/或其分组。
根据上下文，术语“如果”可被解释为意指“当...时”(“when”或“upon”)或“响应于确定”或“响应于检测到”。类似地，根据上下文，短语“如果确定...”或“如果检测到[所陈述的条件或事件]”可被解释为意指“在确定...时”或“响应于确定...”或“在检测到[所陈述的条件或事件]时”或“响应于检测到[所陈述的条件或事件]”。
首先对本申请实施例涉及的若干个名词进行简介：
乳房钼靶图像：是利用X射线的物理性质及人体乳房组织不同的等密度值，将乳房的二维图像投影于X光胶片之上得到的图像，简称钼靶图像。
根据钼靶图像的视图位置不同，包括：CC位的钼靶图像和MLO位的钼靶图像。
图1示出了本申请一个示例性实施例提供的计算机系统的结构框图。该计算机系统包括：乳腺数字放射(Digital Radiography，DR)设备02、乳腺钼靶图像辅助诊断系统10和医生设备04。
乳腺DR设备02是用于基于数字X射线成像技术，采集乳腺钼靶图像的设备。乳腺DR设备02可属于乳腺钼靶图像辅助诊断系统10，也可以独立于乳腺钼靶图像辅助诊断系统10之外部署，图1中以独立部署为例来说明。
乳腺钼靶图像辅助诊断系统10是一台或多台计算机设备，比如单台服务器、服务器群组、服务器集群中的至少一种。系统10包括：乳房信息提取子系统12、病灶识别子系统14、乳房良恶性子系统16和自动化报告生成子系统18中的至少一种。
乳房信息提取子系统12用于提取乳房整体信息。该子系统12包括：乳头位置检测模型、肌肉位置检测模型和腺体类型分类模型。其中，乳头位置检测模型和肌肉位置检测模型可以实现成为同一个神经网络模型，也即乳头和肌肉检测模型。
病灶识别子系统14用于对乳腺钼靶图像进行肿块检测和钙化检测。检测 过程中会考虑乳房的腺体类型进行阈值的选择。病灶识别子系统14对检测到的病灶进行病灶良恶性分析、病灶属性描述、CC位病灶和MLO位病灶匹配。此外，还根据乳头位置和肌肉位置定位病灶所在的象限。
乳房良恶性子系统16用于以病灶良恶性分析子系统14作为基础，结合CC位的钼靶图像和MLO位的钼靶图像进行单侧乳房的良恶性预测。乳房良恶性子系统16包括有乳房良恶性检测模型。可选地，乳房良恶性检测模型在预测时根据乳房的腺体类型进行预测阈值的动态选择。
自动化报告生成子系统18用于对上述三个子系统的所有预测结果进行融合，生成结构化的检测报告。医生可对生成的检测报告进行校正，并可以选择感兴趣的钼靶图像区域，通过交互式查询方式得到该钼靶图像区域的相关信息。
医生设备04是医生使用的计算机设备，可以是台式电脑、笔记本、触摸显示屏等。医生设备04用于查看自动化报告生成子系统18所生成的检测报告以及相应的人机交互。医生设备04可属于乳腺钼靶图像辅助诊断系统10，也可以独立于乳腺钼靶图像辅助诊断系统10之外部署，图1中以独立部署为例来说明。
图2示出了本申请一个示例性实施例提供的乳腺钼靶图像的辅助诊断方法的流程图。该方法可以应用于图1所示的乳腺钼靶图像的辅助诊断系统(以下简称为计算机设备)。该方法包括：
步骤202，获取单侧乳房的乳腺钼靶图像，乳腺钼靶图像包括：CC位钼靶图像和MLO位钼靶图像；
计算机设备获取单侧乳房在两个不同视图位上的乳腺钼靶图像。在一些实施例中，两个不同视图位上的乳腺钼靶图像包括CC位钼靶图像和MLO位钼靶图像，其中：CC位的钼靶图像是采用头尾位采集的钼靶图像。CC位钼靶图像的中心线是自上而下，经乳腺上方垂直入射胶片中心。MLO位的钼靶图像是采用内侧斜位采集的钼靶图像。MLO位钼靶图像的中心线是经乳腺内 侧垂直入射胶片中心。
步骤204，调用乳房良恶性检测模型对CC位钼靶图像和MLO位钼靶图像进行良恶性预测，得到单侧乳房的良恶性预测结果；
在一些实施例中，乳房良恶性检测模型包括：第一单图检测部22、第二单图检测部24、池化层26和全连接层28。
计算机设备调用第一单图检测部22对CC位钼靶图像进行处理，得到第一特征(比如第一logits特征)；调用第二单图检测部24对MLO位钼靶图像进行处理，得到第二特征(比如第二logits特征)。计算机设备将第一特征和第二特征输入池化层26和全连接层28，得到单侧乳房的良恶性预测结果。
步骤206，生成和输出检测报告，检测报告包括单侧乳房的良恶性预测结果。
综上所述，本实施例提供的方法，通过获取单侧乳房的CC位钼靶图像和MLO位钼靶图像；调用乳房良恶性检测模型对CC位钼靶图像和MLO位钼靶图像进行良恶性预测，得到单侧乳房的良恶性预测结果；生成和输出检测报告，检测报告包括单侧乳房的良恶性预测结果；由于乳房良恶性检测模型能够综合对CC位钼靶图像和MLO位钼靶图像进行良恶性预测，也即综合了两个不同视角的钼靶图像进行良恶性预测，较为真实地模拟了医生实际的看片过程，从提高了对单侧乳房的良恶性预测结果的准确性。
在基于图2的可选实施例中，计算机设备调用腺体类型分类模型对乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果，根据腺体类型识别结果确定所述乳房良恶性检测模型对应的预测阈值，调用确定预测阈值后的乳房良恶性检测模型对CC位钼靶图像和MLO位钼靶图像进行良恶性预测，得到单侧乳房的良恶性预测结果。
图4示出了本申请一个示例性实施例提供的乳腺钼靶图像的辅助诊断方法的流程图。该方法可以应用于图1所示的乳腺钼靶图像辅助诊断系统(以下称计算机设备)。该方法包括：
步骤401，获取单侧乳房的乳腺钼靶图像，乳腺钼靶图像包括：CC位钼靶图像和MLO位钼靶图像。
计算机设备从乳腺DR设备获取单侧乳房的乳腺钼靶图像。或者，计算机设备接收医生设备上传的单侧乳房的乳腺钼靶图像。
单侧乳房是指左侧乳房或右侧乳房。当单侧乳房是左侧乳房时，计算机设备获取左侧乳房的乳腺钼靶图像。当单侧乳房是右侧乳房时，计算机设备获取右侧乳房的乳腺钼靶图像。
步骤402，调用腺体类型分类模型对CC位钼靶图像的腺体类型进行识别，得到第一腺体类型。
第一腺体类型包括：脂肪型、少量腺体型、多量腺体型和致密型中的任意一种。
步骤403，调用腺体类型分类模型对MLO位钼靶图像的腺体类型进行识别，得到第二腺体类型。
第二腺体类型包括：脂肪型、少量腺体型、多量腺体型和致密型中的任意一种。
步骤404，将第一腺体类型和第二腺体类型中腺体密度较大的一种，确定为单侧乳房的腺体类型。
在一些实施例中，四种腺体类型的腺体密度由大到小排列的顺序为：致密型＞多量腺体型＞少量腺体型＞脂肪型。
在一些实施例中，第一腺体类型是腺体型，第二腺体类型是少量腺体型，则将腺体型确定为单侧乳房的腺体类型。在另一些示例中，第一腺体类型是致密型，第二腺体类型是脂肪型，则将致密型确定为单侧乳房的腺体类型。
步骤405，根据腺体类型识别结果确定乳房良恶性检测模型对应的预测阈值。
乳房良恶性检测模型中存在用于检测良恶性的概率阈值，计算机设备根据腺体类型识别结果来确定相应的预测阈值。在一些实施例中，脂肪型对应第一阈值，少量腺体型对应第二阈值，多量腺体型对应第三阈值，致密型对 应第四阈值。
当单侧乳房的腺体类型是脂肪型时，计算机设备确定乳房良恶性检测模型对应的预测阈值为第一阈值。当单侧乳房的腺体类型是少量腺体型时，计算机设备确定乳房良恶性检测模型对应的预测阈值为第二阈值。当单侧乳房的腺体类型是多量腺体型时，计算机设备确定乳房良恶性检测模型对应的预测阈值为第三阈值。当单侧乳房的腺体类型是致密型时，计算机设备确定乳房良恶性检测模型对应的预测阈值为第四阈值。
步骤406，调用第一单图检测部对CC位钼靶图像进行处理，得到第一特征。
步骤407，调用第二单图检测部对MLO位钼靶图像进行处理，得到第二特征。
步骤408，将第一特征和第二特征输入池化层和全连接层，得到单侧乳房的良恶性预测结果。
示意性的，以预测阈值为60％为例，当全连接层输出的良恶性预测概率为65％时，由于65％大于60％，则确定良恶性预测结果为恶性。当全连接层输出的良恶性预测概率为40％时，由于40％小于60％，则确定良恶性预测结果为良性。
其中，预测阈值可根据腺体类型而动态改变。
综上所述，本实施例提供的方法，通过获取单侧乳房的CC位钼靶图像和MLO位钼靶图像；调用乳房良恶性检测模型对CC位钼靶图像和MLO位钼靶图像进行良恶性预测，得到单侧乳房的良恶性预测结果；生成和输出检测报告，检测报告包括单侧乳房的良恶性预测结果；由于乳房良恶性检测模型能够综合对CC位钼靶图像和MLO位钼靶图像进行良恶性预测，也即综合了两个不同视角的钼靶图像进行良恶性预测，较为真实地模拟了医生实际的看片过程，从提高了对单侧乳房的良恶性预测结果的准确性。
下文对上述各个神经网络模型以及相应的神经网络模型的训练方法进行 介绍。
针对乳房信息提取子系统
乳房信息提取子系统包括：乳头检测模型、肌肉检测模型和腺体类型分类模型。其中，乳头检测模型和肌肉检测模型可设计为同一个检测模型：乳头和肌肉检测模型，如图5所示。
在一些实施例中，乳头检测模型是基于全卷积网络(Fully Convolutional Networks，FCN)的两分类(乳头+背景)模型。肌肉检测模型是基于FCN的两分类(肌肉+背景)模型。乳头和肌肉检测模型是基于FCN的三分类(乳头+肌肉+背景)。
在基于图2至图4任一所示的可选实施例中，还包括如下步骤：
计算机设备调用乳头检测模型(或乳头和肌肉检测模型)对乳腺钼靶图像中的乳头位置进行识别。乳头检测模型用于对乳腺钼靶图像中的各个像素点进行标定，每个像素点的类型可以是乳头和背景中的一种。
计算机设备调用肌肉检测模型(或乳头和肌肉检测模型)对乳腺钼靶图形象中的肌肉位置进行识别。肌肉检测模型用于对乳腺钼靶图像中的各个像素点进行标定，每个像素点的类型可以是肌肉和背景中的一种。
计算机设备调用乳头和肌肉检测模型对乳腺钼靶图像中的乳头位置和肌肉位置进行识别。乳头和肌肉检测模型用于对乳腺钼靶图像中的各个像素点进行标定，每个像素点的类型可以是乳头、肌肉和背景中的一种。
对于CC位钼靶图像，乳头和肌肉检测模型用于检测乳头区域；对于MLO位钼靶图像，乳头和肌肉检测模型用于检测乳头和肌肉区域。
上述乳头和肌肉检测模型的训练方法可以如下，如图6所示：
步骤601，对乳头和肌肉检测模型进行权重初始化。
在一些实施例中，计算机设备可采用加州大学伯克利分校(UC Berkeley)发布的FCN模型构建乳头和肌肉检测模型，采用分割数据集(pattern analysis,statistical modelling and computational learning visual object classes，PASCAL  VOC)对乳头和肌肉检测模型的权重进行初始化。PASCAL VOC数据集是一种提供用于对象类识别的标准化图像数据集，也可以为提供用于访问数据集和注释的公共工具集。
步骤602，获取训练样本，训练样本包括样本乳腺钼靶图像、乳头标定结果和肌肉标定结果。
在一些实施例中，训练样本包括两部分：医学钼靶图像公开数据集(Digital Database for Screening Mammography，DDSM)数据库和手工标定数据集。手工标定数据集可以是使用国内医院数据，聘请专家进行像素级标准后的样本乳腺钼靶图像(1000张+)。可选地，对于手工标定的数据集，还可以采用图像翻转和/或图像裁剪技术进行数据增强。
其中，DDSM数据库是医学机构建立的专门存放乳腺癌图像的数据库。DDSM数据库里存放了恶性、常规、良性等数据类型，目前很多对乳腺癌的研究都是依据DDSM数据库进行研究。
第一训练过程可采用公开数据集DDSM中的样本乳腺钼靶图像进行训练，然后再采用手工标定数据集进行迁移学习。迁移学习的参数可以为：样本乳腺钼靶图像的输入尺寸为800*800像素、批处理大小为8，学习率为0.00001，最大迭代次数10000)。
其中，迁移学习是把已学训练好的模型参数迁移到新的模型来帮助新模型训练。考虑到大部分数据或任务是存在相关性的，所以通过迁移学习可以将已经学到的模型参数通过某种方式来分享给新模型从而加快并优化模型的学习效率不用像大多数网络那样从零学习。在本申请中，通过参数迁移的方式进行网络模型数据训练，能够将任务A训练出来的模型用来初始化任务B的模型参数，使任务B能够更快的学习训练收敛。
步骤603，使用乳头和肌肉检测模型对训练样本进行预测，得到预测结果。
步骤604，根据预测结果、乳头标定结果和肌肉标定结果进行误差计算，得到误差损失。
步骤605，根据误差损失采用误差反向传播算法对乳头和肌肉检测模型进行训练。
当训练次数达到预设次数，或误差损失小于阈值时，认为满足训练结束条件。训练得到用于提取乳头区域和肌肉区域的全卷积分割网络。
腺体类型分类模型可以是以谷歌公司发布的Inception V3模型来构建的分类模型。腺体类型包括：脂肪型、少量腺体型、多量腺体型和致密型四种中的一种。其中，Inception V3模型是卷积神经网络中的一种。卷积神经网络是一种前馈神经网络，人工神经元可以响应周围单元，可以进行大型图像处理，卷积神经网络包括卷积层和池化层。Inception V3模型通过增加单层卷积层的宽度，即在单层卷积层上使用不同尺度的卷积核，从而优化网络。Inception V3模型通过密集成分来近似最优的局部稀疏结，从而更高效的利用计算资源，在相同的计算量下能提取到更多的特征，从而提升训练结果。Inception V3模型的特点有两个：一是使用1x1的卷积来进行升降维；二是在多个尺寸上同时进行卷积再聚合。
其中，卷积层由若干卷积单元组成，每个卷积单元的参数都是通过反向传播算法最佳化得到的。图像卷积运算的目的是提取输入图像的不同特征，第一层卷积层可能只能提取一些低级的特征如边缘、线条和角等层级，更多层的网络能从低级特征中迭代提取出更复杂的特征。
上述腺体类型分类模型的训练方法可以如下，如图7所示：
步骤701，对腺体类型分类模型进行权重初始化。
在一些实施例中，计算机设备可采用谷歌公司发布的Inception V3模型构建腺体类型分类模型，输出的分类类别设置为4。然后使用ImageNet(计算机视觉标准数据集)数据集进行权重初始化。
步骤702，获取训练样本，训练样本包括样本乳腺钼靶图像和腺体类型标定结果。
在一些实施例中，训练样本包括两部分：谷歌公司发布公开数据集DDSM 和手工标定数据集。手工标定数据集可以是使用国内医院数据，聘请专家进行腺体类型进行标定后的样本乳腺钼靶图像(1000张+)。可选地，对于手工标定的数据集，还可以采用图像翻转和/或图像裁剪技术进行数据增强。
第一训练过程可采用公开数据集DDSM中的样本乳腺钼靶图像进行训练，然后再采用手工标定数据集进行迁移学习。迁移学习的参数可以为：误差反向传播算法使用均方根误差(Root Mean Square prop，RMSprop)，批处理大小为64，初始学习率为0.00001，最大迭代次数为10000。
步骤703，使用腺体类型分类模型对训练样本进行预测，得到预测结果。
步骤704，根据预测结果和腺体类型标定结果进行误差计算，得到误差损失。
步骤705，根据误差损失采用误差反向传播算法对腺体类型分类模型进行训练。
当训练次数达到预设次数，或误差损失小于阈值时，认为满足训练结束条件。训练得到用于识别腺体类型的属性分类网络。
针对病灶识别子系统
病灶识别子系统包括：病灶描述模型、病灶良恶性模型、病灶匹配模型和病灶象限定位模型，如图8所示。
在基于图2至图4的可选实施例中，还包括如下四个步骤中的至少一个步骤：
计算机设备调用病灶描述模型对乳腺钼靶图像中的病灶进行检测，得到病灶描述信息。
计算机设备调用病灶良恶性模型对乳腺钼靶图像中的病灶进行良恶性识别，得到病灶良恶性概率。
计算机设备调用病灶匹配模型对CC位钼靶图像和MLO位钼靶图像的病灶进行一致性判断，得到病灶匹配概率。
计算机设备调用病灶象限定位模型对乳腺钼靶图像中的病灶进行象限计 算。
病灶描述模型可以是以谷歌公司发布的Inception V3模型来构建的分类模型。上述病灶描述模型的训练方法可以如下，如图9所示：
步骤901，对病灶描述模型进行权重初始化。
在一些实施例中，采用谷歌公司发布的Inception V3模型构建病灶描述模型。在一些实施例中，将Inception V3模型的最后一个全连接层修改为支持多个任务同时训练的多个并列的全连接层，每个任务对应的输出类别设置为2，也即每个任务共享除最后一个全连接层之外的所有参数。然后使用ImageNet数据集进行权重初始化。
步骤902，获取训练样本，训练样本包括样本乳腺钼靶图像和病灶标定结果。
在一些实施例中，训练样本包括两部分：谷歌公司发布公开数据集DDSM和手工标定数据集。手工标定数据集可以是使用国内医院数据，聘请专家对病灶属性进行标定后的样本乳腺钼靶图像(1000张+)。以肿块属性为例，对每一个肿块标注圆形或不规则形、边界清晰或模糊、边界见分页或不见分页、边界见毛刺或不见毛刺中的至少一项描述信息。可选地，对于手工标定的数据集，还可以采用图像翻转和/或图像裁剪技术进行数据增强。
第一训练过程可采用公开数据集DDSM中的样本乳腺钼靶图像进行训练，然后再采用手工标定数据集进行迁移学习。迁移学习的参数可以为：误差反向传播算法使用Adam，批处理大小为64，初始学习率为0.001，最大迭代次数为10000。
步骤903，使用病灶描述模型对训练样本进行预测，得到预测结果。
步骤904，根据预测结果和病灶标定结果进行误差计算，得到误差损失。
步骤905，根据误差损失采用误差反向传播算法对病灶描述模型进行训练。
当训练次数达到预设次数，或误差损失小于阈值时，认为满足训练结束 条件。训练得到用于识别病灶的属性分类网络。
病灶良恶性模型可以是以谷歌公司发布的Inception V3模型来构建的分类模型。上述病灶良恶性模型的训练方法可以如下，如图10所述：
步骤1001，对病灶良恶性模型进行权重初始化。
在一些实施例中，采用谷歌公司发布的Inception V3模型构建病灶良恶性模型。在一些实施例中，将Inception V3模型的最后一个池化层修改为最大池化层(max pooling)，输出类别数设置为2。然后使用ImageNet数据集进行权重初始化。
步骤1002，获取训练样本，训练样本包括样本乳腺钼靶图像和良恶性标定结果。
在一些实施例中，训练样本包括两部分：谷歌公司发布公开数据集DDSM和手工标定数据集。手工标定数据集可以是使用国内医院数据，聘请专家对病灶的良恶性进行标定后的样本乳腺钼靶图像(16000张+)。其中，良恶性标定结果包括：恶性钙化病灶和恶性肿块病灶为正样本，良性钙化病灶、良性肿块病灶和正常区域为负样本。可选地，对于手工标定的数据集，还可以采用图像翻转和/或图像裁剪技术进行数据增强。
第一训练过程可采用公开数据集DDSM中的样本乳腺钼靶图像进行训练，然后再采用手工标定数据集进行迁移学习。迁移学习的参数可以为：误差反向传播算法使用Adam，批处理大小为64，初始学习率为0.001，最大迭代次数为10000。
步骤1003，使用病灶良恶性模型对训练样本进行预测，得到预测结果。
步骤1004，根据预测结果和病灶标定结果进行误差计算，得到误差损失。
步骤1005，根据误差损失采用误差反向传播算法对病灶良恶性模型进行训练。
当训练次数达到预设次数(比如10000)，或误差损失小于阈值时，认为满足训练结束条件。训练得到用于识别病灶的属性分类网络。在一些实施例 中，概率大于0.5被认为是疑似包含恶性病灶。
病灶象限定位模型是基于肌肉直线拟合方程的模型，通过对乳腺钼靶图像中的肌肉位置的边界进行线性拟合，得到肌肉边界直线方程。然后，根据乳头位置对乳腺钼靶图像中的病灶进行象限计算。
计算机设备获取病灶象限定位模型，其中病灶象限定位模型是通过多个样本图像训练得到的，样本图像中的每一个像素点都进行了标注，标注类型包括：背景、乳头和肌肉，换句话说，病灶象限定位模型可以识别出一张图像中每个像素点要么是属于背景的，要么是属于乳头的要么是属于肌肉的。计算机设备将CC位钼靶图像输入病灶象限定位模型，基于病灶象限定位模型可以确定CC位钼靶图像中的乳头位置(CC位钼靶图像是没有肌肉信息的，因此没有肌肉区域)；计算机设备将MLO位钼靶图像输入病灶象限定位模型，基于病灶象限定位模型可以确定MLO位钼靶图像中的乳头位置以及肌肉位置。对CC位钼靶图像来说，计算机设备根据乳头位置以及乳腺边缘分界线，确定第一分割线，根据第一分割线，确定第一病灶区域位于内象限或者外象限。对MLO位钼靶图像来说，计算机设备根据肌肉位置拟合肌肉边界线直线方程，进而确定肌肉边界线(即是前述中的对象分界线)，再根据乳头位置以及肌肉边界线，确定第二分割线，根据第二分割线，确定第二病灶区域位于上象限或者下象限。
结合图11所示出的用户界面20a所示，计算机设备10a获取同一个患者的同一侧乳房的乳腺钼靶图像，并将获取到的乳腺钼靶图像显示在屏幕上，其中乳腺钼靶图像包括：CC位钼靶图像20b，以及MLO位钼靶图像20c，CC位钼靶图像是按照头尾位对乳腺成像，MLO位钼靶图像是按照斜侧位对乳腺成像。
计算机设备10a获取肿块检测模型以及钙化检测模型(也即本申请中的病灶描述模型)，肿块检测模型可以识别出图像中肿块病灶所在的位置区域；钙化检测模型可以是识别出图像中钙化病灶所在的位置区域，肿块类别和钙 化类别属于病灶类别。
对肿块检测模型来说，计算机设备10a可以将CC位钼靶图像20b输入肿块检测模型，肿块检测模型可以输出CC位钼靶图像20b中的病灶对象位于CC位钼靶图像20b中的病灶区域20d，还可以确定CC位钼靶图像20b中的病灶对象所属的病灶类别为肿块类别。
计算机设备10a可以将MLO位钼靶图像20c输入肿块检测模型，肿块检测模型也可以输出MLO位钼靶图像20c中的病灶对象位于MLO位钼靶图像20c中的病灶区域20h，且还可以确定MLO位钼靶图像20c中的病灶对象所属的病灶类别为肿块类别。
对钙化检测模型来说，计算机设备10a同样将CC位钼靶图像20b输入钙化检测模型，钙化检测模型在CC位钼靶图像20b中没有检测到钙化病灶；计算机设备10a同样将MLO位钼靶图像20c输入钙化检测模型，钙化检测模型在MLO位钼靶图像20c中同样没有检测到钙化病灶。
因此，对CC位钼靶图像20b，以及MLO位钼靶图像20c来说，只存在肿块病灶，且在CC位钼靶图像20b中肿块病灶位于病灶区域20d；在MLO位钼靶图像20c中肿块病灶位于病灶区域20h。
计算机设备10a获取病灶象限定位模型，病灶象限定位模型可以是识别图像中每个像素点所属的组织类别，组织类别包括：乳头类别、肌肉类别以及背景类别。
计算机设备10a将CC位钼靶图像20b输入病灶象限定位模型，模型可以确定CC位钼靶图像20b的每个像素点所属的组织类别。在CC位钼靶图像20b中，计算机设备10a将属于乳头类别的像素组合为区域20e，该区域20e即是乳头所在的区域。计算机设备10a在CC位钼靶图像20b中确定乳腺边缘线20g，将垂直于乳腺边缘线20g且经过区域20e的线条20f作为象限分割线20f。在CC位钼靶图像20b中，位于象限分割线20f以下的为内象限，位于象限分割线20f以上的为外象限。在CC位钼靶图像20b中，由于病灶区域20d位于内象限(病灶区域20d中的大部分位于内象限，那就认为病灶区域 20d位于内象限)，因此计算机设备10a可以确定CC位钼靶图像20b中的肿块病灶位于内象限。
计算机设备10a将MLO位钼靶图像20c输入病灶象限定位模型，模型可以确定MLO位钼靶图像20c的每个像素点所属的组织类别。在MLO位钼靶图像20c中，计算机设备10a将属于乳头类别的像素组合为区域20j，该区域20j即是乳头所在的区域。计算机设备10a将属于肌肉类别的象限组合为肌肉区域，并确定肌肉区域与非肌肉区域的区域分界线20m，将垂直于区域分界线20m且经过区域20j的线条20k作为象限分割线20k。在MLO位钼靶图像20c中，位于象限分割线20k以下的为下象限，位于象限分割线20k以上的为上象限。在MLO位钼靶图像20c中，由于病灶区域20h位于下象限，因此计算机设备10a可以确定MLO位钼靶图像20c中的肿块病灶位于下象限。
计算机设备10a将由CC位钼靶图像20b确定的内象限，以及由MLO位钼靶图像20c确定的下象限，组合为象限位置信息20n“内下象限”。
计算机设备10a可以将象限位置信息20n“内下象限”以及与CC位钼靶图像20b、MLO位钼靶图像20c均对应的病灶类别“肿块”，组合为诊断意见：“内下象限见肿块”。
病灶匹配模型可以是基于牛津大学发布的VGG模型构建的分类模型。病灶匹配模型的训练过程可以如下，如图12所示：
步骤1201，对病灶匹配模型进行权重初始化。
在一些实施例中，采用牛津大学发布的VGG模型构建病灶匹配模型。在一些实施例中，取VGG模型的pool5层结果进行融合，后续与原始VGG一致采用三个全连接层得到分类数为2个结果。然后使用ImageNet数据集进行权重初始化。
步骤1202，获取训练样本，训练样本包括样本乳腺钼靶图像和匹配标定结果。
在一些实施例中，训练样本包括两部分：谷歌公司发布公开数据集DDSM 和手工标定数据集。手工标定数据集可以是使用国内医院数据，聘请专家对CC位钼靶图像和MLO位钼靶图像中表示同一个病灶的一对斑块作为正样本，其他任意两个不表示同一个病灶的一对斑块作为负样本，得到样本乳腺钼靶图像(1000张+)。可选地，对于手工标定的数据集，还可以采用图像翻转和/或图像裁剪技术进行数据增强。
第一训练过程可采用公开数据集DDSM中的样本乳腺钼靶图像进行训练，然后再采用手工标定数据集进行迁移学习。迁移学习的参数可以为：误差反向传播算法使用Adam，批处理大小为128，初始学习率为0.001，最大迭代次数为10000。
步骤1203，使用病灶匹配模型对训练样本进行预测，得到预测结果。
步骤1204，根据预测结果和匹配标定结果进行误差计算，得到误差损失。
步骤1205，根据误差损失采用误差反向传播算法对病灶匹配模型进行训练。
当训练次数达到预设次数(比如10000)，或误差损失小于阈值时，认为满足训练结束条件。训练得到用于识别病灶匹配程度的属性分类网络。在一些实施例中在预测阶段中，同时输入CC位的钼靶图像中的第一病灶和MLO位的钼靶图像中的第二病灶，得到第一病灶和第二病灶是否属于同一病灶的概率。
针对乳房良恶性子系统
乳房良恶性子系统中包括：乳房良恶性模型，如图13所示。乳房良恶性模型用于对单侧乳房的CC位钼靶图像和MLO位钼靶图像进行良恶性检测，得到单侧乳房的良恶性概率，乳房良恶性模型的结构如图3所示。
在基于图2至图4的可选实施例中，上述乳房良恶性模型的训练过程包括：单图良恶性模型的训练，以及乳房良恶性模型的训练。单图良恶性模型用于构建图3中的第一单图检测部22和第二单图检测部24。
单图良恶性模型的训练方法包括如下：
步骤1401，获取训练好的病灶良恶性模型作为初始的单图良恶性模型。
步骤1402，获取训练样本，训练样本包括样本乳腺钼靶图像和整图良恶性标定结果；
在一些实施例中，训练样本包括两部分：手工标定数据集。手工标定数据集可以是使用国内医院数据，聘请专家对整张图片的良恶性进行标定后的样本乳腺钼靶图像(16000张+)。其中，整图良恶性标定结果包括：恶性的钼靶图像作为正样本，整图为良性和/或正常的钼靶图像作为负样本。可选地，对于手工标定的数据集，还可以采用图像翻转和/或图像裁剪技术进行数据增强。
采用手工标定数据集对初始的单图良恶性模型进行迁移学习。迁移学习的参数可以为：误差反向传播算法使用Adam，批处理大小为64，初始学习率为0.001，最大迭代次数为10000。
步骤1403，使用单图良恶性模型对训练样本进行预测，得到预测结果。
步骤1404，根据预测结果和整图良恶性标定结果进行误差计算，得到误差损失。
步骤1405，根据误差损失采用误差反向传播算法对单图良恶性模型进行训练。
当训练次数达到预设次数(比如10000)，或误差损失小于阈值时，认为满足训练结束条件。训练得到用于识别单图良恶性的属性分类网络。在一些实施例中，概率大于0.5被认为是疑似包含恶性病灶。
根据训练完毕的两个单图良恶性模型分别作为第一单图检测部22和第二单图检测部24，通过增加池化层25和全连接层28，使模型的输出类别保持为2后，形成乳房良恶性模型。
使用单图良恶性模型的训练样本作为新的训练样本，通过数据增强之后(由于是钼靶图片，主要进行翻转和裁剪的数据增强，无须进行颜色空间的数据增强)，通过训练将单图良恶性分类模型扩展为CC和MLO双图良恶性 分类模型。训练参数可以为：(下降算法使用RMSprop，批处理大小为32，初始学习率为0.01，最大迭代次数为10000)。
针对自动报告生成子系统
图15(A)示出了自动报告生成子系统将上述子系统的结果进行融合并自动生成结构化报告的原理示意图。自动报告生成子系统通过对上述各子系统的所有检测与识别结果进行融合，生成结构化报告，医生可以对生成的报告进行修改，并可以通过交互式方式，查询感兴趣区域的相关信息。各步骤的详细说明如下：
自动报告生成子系统具有如下功能：
结构化的检测报告生成
计算机设备通过对上述所有子系统得到的检测与识别结果进行汇总，自动生成BI-RADS标准中所述的检测报告的报告内容。检测报告的报告内容包括：肿块描述、钙化描述、腺体类型描述以及乳房良恶性的描述等，在一些实施例中，该检测报告可参考图15(B)所示。检测报告包括：少量腺体型、恶性肿块、恶性钙化、右侧乳房恶性概率较大、内下象限肿块以及肿块的描述信息等，其中肿块的描述信息包括：不规则形、边界不清、带毛刺、及浅分页中的至少一种。
检测报告的校正：
医生可以通过复核等方式，对本系统的检测与识别结果和自动生成的报告进行修正，得到诊断报告。计算机设备接收医生设备发送的报告校正请求，根据报告校正请求对检测报告进行校正。
交互式查询：
医生可以通过交互式方式，查询其感兴趣的信息。例如，医生可选择钼靶图像中感兴趣的区域，通过调用本系统的病灶良恶性分类模型，得到该区域的良恶性结果。计算机设备还接收局部查询请求，该局部查询请求用于请求查询乳腺钼靶图像中的局部区域，根据局部查询请求输出局部区域对应的 检测报告。
参考图16，本申请还提供了一种乳腺钼靶图像处理方法，该方法由计算机设备执行，该乳腺钼靶图像处理方法包括以下步骤：
步骤1602，获取单侧乳房的乳腺钼靶图像，该钼靶图像包括：头尾CC位钼靶图像和内侧斜MLO位钼靶图像。
步骤1604，调用目标检测模型对该CC位钼靶图像和该MLO位钼靶图像进行处理，得到该单侧乳房的图像检测结果。
步骤1606，根据该图像检测结果生成检测报告并输出。
在一些实施例中，该目标检测模型包括：第一单图检测部、第二单图检测部、池化层和全连接层。调用目标检测模型对该CC位钼靶图像和该MLO位钼靶图像进行处理，得到该单侧乳房的图像检测结果的步骤具体包括：调用该第一单图检测部对该CC位钼靶图像进行处理，得到第一特征；调用该第二单图检测部对该MLO位钼靶图像进行处理，得到第二特征；及将该第一特征和该第二特征输入该池化层和该全连接层，得到该单侧乳房的图像检测结果。
在一些实施例中，调用目标检测模型对该CC位钼靶图像和该MLO位钼靶图像进行处理，得到该单侧乳房的图像检测结果之前，该乳腺钼靶图像处理方法还包括：调用腺体类型分类模型对该乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果。调用目标检测模型对该CC位钼靶图像和该MLO位钼靶图像进行处理，得到该单侧乳房的图像检测结果的步骤具体包括：根据该腺体类型识别结果确定该目标检测模型对应的预测阈值；及调用确定预测阈值后的目标检测模型对该CC位钼靶图像和该MLO位钼靶图像进行处理，得到该单侧乳房的图像检测结果。
在一些实施例中，调用腺体类型分类模型对该乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果的步骤具体包括：调用该腺体类型分类模型对该CC位钼靶图像的腺体类型进行识别，得到第一腺体类型；调用该 腺体类型分类模型对该MLO位钼靶图像的腺体类型进行识别，得到第二腺体类型；及将该第一腺体类型和该第二腺体类型中腺体密度较大的一种，确定为该单侧乳房的腺体类型。
在一些实施例中，该乳腺钼靶图像处理方法还包括：调用异常识别模型对该乳腺钼靶图像进行异常检测，得到异常检测结果，该异常检测包括肿块检测和钙化检测中的至少一种。
在一些实施例中，该异常识别模型包括：异常描述模型、异常分类模型、异常匹配模型和异常象限定位模型中的至少一种；该调用异常识别模型对该乳腺钼靶图像进行异常检测，得到异常检测结果的步骤，具体包括以下步骤中的至少一种：调用该异常描述模型对该乳腺钼靶图像中的异常区域进行检测，得到异常描述信息；调用该异常分类模型对该乳腺钼靶图像中的异常类别进行识别，得到对应的类别概率；调用该异常匹配模型对该CC位钼靶图像和该MLO位钼靶图像的异常区域进行一致性判断，得到异常区域匹配概率；及调用该异常象限定位模型对该乳腺钼靶图像中的异常区域进行象限计算。
在一些实施例中，调用该异常象限定位模型对该乳腺钼靶图像中的异常区域进行象限计算之前，该乳腺钼靶图像处理方法还包括：调用乳头检测模型对该乳腺钼靶图像中的乳头位置进行识别。调用该异常象限定位模型对该乳腺钼靶图像中的异常区域进行象限计算的步骤具体包括：调用该异常象限定位模型根据该乳头位置对该乳腺钼靶图像中的异常区域进行象限计算。
在一些实施例中，该乳腺钼靶图像处理方法还包括：当该乳腺钼靶图像是MLO位钼靶图像时，调用肌肉检测模型对该乳腺钼靶图像中的肌肉位置进行识别。
在一些实施例中，该乳腺钼靶图像处理方法还包括：接收报告校正请求；及根据该报告校正请求对该检测报告进行校正。
在一些实施例中，该乳腺钼靶图像处理方法还包括：接收局部查询请求，该局部查询请求用于请求查询该乳腺钼靶图像中的局部区域；及根据该局部 查询请求输出该局部区域对应的检测报告。
需要说明的是，关于该乳腺钼靶图像处理方法的具体实施细节，可参考前述实施例所提及的乳腺钼靶图像的辅助诊断方法的描述内容。其中，该目标检测模型的具体处理过程可参考前述实施例所提及的乳腺钼靶图像的辅助诊断方法中乳房良恶性检测模型的处理过程；该异常识别模型的具体处理过程可参考前述实施例所提及的乳腺钼靶图像的辅助诊断方法中病灶识别模型的处理过程；该异常描述模型的具体处理过程可参考前述实施例所提及的乳腺钼靶图像的辅助诊断方法中病灶描述模型的处理过程；该异常分类模型的具体处理过程可参考前述实施例所提及的乳腺钼靶图像的辅助诊断方法中病灶良恶性模型的处理过程；该异常匹配模型的具体处理过程可参考前述实施例所提及的乳腺钼靶图像的辅助诊断方法中病灶匹配模型的处理过程；该异常象限定位模型的具体处理过程可参考前述实施例所提及的乳腺钼靶图像的辅助诊断方法中病灶象限定位模型的处理过程。
图17示出了本申请一个示例性实施例提供的乳腺钼靶图像的辅助诊断装置的框图。该装置可以用于实现上述乳腺钼靶图像的辅助诊断方法的功能。该装置包括：
图像获取模块101，用于获取单侧乳房的乳腺钼靶图像，该乳腺钼靶图像包括：CC位钼靶图像和MLO位钼靶图像。
乳房良恶性检测模型102，用于对该CC位钼靶图像和该MLO位钼靶图像进行良恶性预测，得到该单侧乳房的良恶性预测结果。
自动化报告生成模块103，用于生成和输出检测报告，该检测报告包括该单侧乳房的良恶性预测结果。
在一些实施例中，该乳房良恶性检测模型102包括：第一单图检测部、第二单图检测部、池化层和全连接层；该乳房良恶性检测模型102，用于调用该第一单图检测部对该CC位钼靶图像进行处理，得到第一logits特征。该乳房良恶性检测模型102，用于调用该第二单图检测部对该MLO位钼靶图像 进行处理，得到第二logits特征。该乳房良恶性检测模型102，用于将该第一logits特征和该第二logits特征输入该池化层和该全连接层，得到该单侧乳房的良恶性预测结果。
在一些实施例中，该乳腺钼靶图像的辅助诊断装置还包括：腺体类型分类模型。该腺体类型分类模型，用于对该乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果。该乳房良恶性检测模型102，用于根据该腺体类型识别结果确定该乳房良恶性检测模型对应的预测阈值；调用确定该预测阈值后的乳房良恶性检测模型对该CC位钼靶图像和该MLO位钼靶图像进行良恶性预测，得到该单侧乳房的良恶性预测结果。
在一些实施例中，该腺体类型分类模型，用于调用该腺体类型分类模型对该CC位钼靶图像的腺体类型进行识别，得到第一腺体类型；调用该腺体类型分类模型对该MLO位钼靶图像的腺体类型进行识别，得到第二腺体类型；将该第一腺体类型和该第二腺体类型中腺体密度较大的一种，确定为该单侧乳房的腺体类型。
在一些实施例中，该乳腺钼靶图像的辅助诊断装置还包括：病灶识别模型；该病灶识别模型，用于对该乳腺钼靶图像进行病灶检测，得到病灶检测结果，该病灶检测包括肿块检测和钙化检测中的至少一种。
在一些实施例中，该病灶识别模型包括：病灶描述模型、病灶良恶性模型、病灶匹配模型和病灶象限定位模型中的至少一种；
该病灶描述模型，用于对该乳腺钼靶图像中的病灶进行检测，得到病灶描述信息。
该病灶良恶性模型，用于对该乳腺钼靶图像中的病灶进行良恶性识别，得到病灶良恶性概率。
该病灶匹配模型，用于对该CC位钼靶图像和该MLO位钼靶图像的病灶进行一致性判断，得到病灶匹配概率。
该病灶象限定位模型，用于对该乳腺钼靶图像中的病灶进行象限计算。
在一些实施例中，该乳腺钼靶图像的辅助诊断装置还包括：乳头检测模 型；该乳头检测模型，用于对该乳腺钼靶图像中的乳头位置进行识别。该病灶象限定位模型，用于根据该乳头位置对该乳腺钼靶图像中的病灶进行象限计算。
在一些实施例中，该乳腺钼靶图像的辅助诊断装置还包括：肌肉检测模型；该肌肉检测模型，用于当该乳腺钼靶图像是MLO位钼靶图像时，对该乳腺钼靶图像中的肌肉位置进行识别。
在一些实施例中，该自动化报告生成模块103，用于接收报告校正请求；根据该报告校正请求对该检测报告进行校正。
在一些实施例中，该自动化报告生成模块103，用于接收局部查询请求，该局部查询请求用于请求查询该乳腺钼靶图像中的局部区域；根据该局部查询请求输出该局部区域对应的检测报告。
本申请还提供了一种乳腺钼靶图像处理装置，该乳腺钼靶图像处理装置可以用于实现上述乳腺钼靶图像处理方法的功能。该乳腺钼靶图像处理装置包括：
图像获取模块，用于获取单侧乳房的乳腺钼靶图像，该乳腺钼靶图像包括：CC位钼靶图像和MLO位钼靶图像。
目标检测模型，用于对该CC位钼靶图像和该MLO位钼靶图像进行处理，得到该单侧乳房的图像检测结果。
自动化报告输出模块，用于根据该图像检测结果生成检测报告并输出。
在一些实施例中，目标检测模型包括：第一单图检测部、第二单图检测部、池化层和全连接层。模型检测模型，用于调用第一单图检测部对该CC位钼靶图像进行处理，得到第一logits特征。目标检测模型，用于调用该第二单图检测部对该MLO位钼靶图像进行处理，得到第二logits特征。乳房良恶性检测模型，用于将该第一logits特征和该第二logits特征输入该池化层和该全连接层，得到单侧乳房的图像检测结果。
在一些实施例中，该乳腺钼靶图像处理装置还包括：腺体类型分类模型。 该腺体类型分类模型，用于对该乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果。该目标检测模型，用于根据该腺体类型识别结果确定该目标检测模型对应的预测阈值；调用确定该预测阈值后的目标检测模型对该CC位钼靶图像和该MLO位钼靶图像进行处理，得到该单侧乳房的图像检测结果。
在一些实施例中，该腺体类型分类模型，用于调用该腺体类型分类模型对该CC位钼靶图像的腺体类型进行识别，得到第一腺体类型；调用该腺体类型分类模型对该MLO位钼靶图像的腺体类型进行识别，得到第二腺体类型；将该第一腺体类型和该第二腺体类型中腺体密度较大的一种，确定为该单侧乳房的腺体类型。
在一些实施例中，该乳腺钼靶图像处理装置还包括：异常识别模型。该异常识别模型，用于对该乳腺钼靶图像进行异常检测，得到异常检测结果，该异常检测包括肿块检测和钙化检测中的至少一种。
在一些实施例中，该异常识别模型包括：异常描述模型、异常分类模型、异常匹配模型和异常象限定位模型中的至少一种。
该异常描述模型，用于对该乳腺钼靶图像中的异常区域进行检测，得到异常描述信息。
该异常分类模型，用于对该乳腺钼靶图像中的异常类别进行识别，得到类别概率。
该异常匹配模型，用于对该CC位钼靶图像和该MLO位钼靶图像的异常区域进行一致性判断，得到异常区域匹配概率。
该异常象限定位模型，用于对该乳腺钼靶图像中的异常区域进行象限计算。
在一些实施例中，该乳腺钼靶图像处理装置还包括：乳头检测模型。该乳头检测模型，用于对该乳腺钼靶图像中的乳头位置进行识别。该异常象限定位模型，用于根据该乳头位置对该乳腺钼靶图像中的异常区域进行象限计算。
在一些实施例中，该乳腺钼靶图像处理装置还包括：肌肉检测模型。该肌肉检测模型，用于当该乳腺钼靶图像是MLO位钼靶图像时，对该乳腺钼靶图像中的肌肉位置进行识别。
在一些实施例中，该自动化报告生成模块103，用于接收报告校正请求；根据该报告校正请求对该检测报告进行校正。
在一些实施例中，该自动化报告生成模块103，用于接收局部查询请求，该局部查询请求用于请求查询该乳腺钼靶图像中的局部区域；根据该局部查询请求输出该局部区域对应的检测报告。
图18示出了本申请一个示例性实施例提供的计算机设备的结构示意图。在一些实施例中，计算机设备1800包括中央处理单元(Central Processing Unit，简称：CPU)1801、包括随机存取存储器(random access memory，简称：RAM)1802和只读存储器(read-only memory，简称：ROM)1803的系统存储器1804，以及连接系统存储器1804和中央处理单元1801的系统总线1805。所述计算机设备1800还包括帮助计算机内的各个器件之间传输信息的基本输入/输出系统(I/O系统)1806，和用于存储操作系统1813、客户端1814和其他程序模块1815的大容量存储设备1807。
所述基本输入/输出系统1806包括有用于显示信息的显示器1808和用于用户输入信息的诸如鼠标、键盘之类的输入设备1809。其中所述显示器1808和输入设备1809都通过连接到系统总线1805的输入/输出控制器1180连接到中央处理单元1801。所述基本输入/输出系统1806还可以包括输入/输出控制器1180以用于接收和处理来自键盘、鼠标、或电子触控笔等多个其他设备的输入。类似地，输入/输出控制器1180还提供输出到显示屏、打印机或其他类型的输出设备。
所述大容量存储设备1807通过连接到系统总线1805的大容量存储控制器(未示出)连接到中央处理单元1801。所述大容量存储设备1807及其相关联的计算机可读介质为计算机设备1800提供非易失性存储。也就是说，所 述大容量存储设备1807可以包括诸如硬盘或者只读光盘(Compact Disc Read-Only Memory，简称：CD-ROM)驱动器之类的计算机可读介质(未示出)。
不失一般性，所述计算机可读介质可以包括计算机存储介质和通信介质。计算机存储介质包括以用于存储诸如计算机可读指令、数据结构、程序模块或其他数据等信息的任何方法或技术实现的易失性和非易失性、可移动和不可移动介质。计算机存储介质包括RAM、ROM、可擦除可编程只读存储器(erasable programmable read-only memory，简称：EPROM)、电可擦除可编程只读存储器(electrically erasable programmable read-only memory，简称：EEPROM)、闪存或其他固态存储其技术，CD-ROM、数字通用光盘(Digital Versatile Disc，简称：DVD)或其他光学存储、磁带盒、磁带、磁盘存储或其他磁性存储设备。当然，本领域技术人员可知所述计算机存储介质不局限于上述几种。上述的系统存储器1804和大容量存储设备1807可以统称为存储器。
根据本申请的各种实施例，所述计算机设备1800还可以通过诸如因特网等网络连接到网络上的远程计算机运行。也即计算机设备1800可以通过连接在所述系统总线1805上的网络接口单元1811连接到网络1812，或者说，也可以使用网络接口单元1811来连接到其他类型的网络或远程计算机系统(未示出)。
在一些实施例中，还提供了一种计算机设备，包括存储器和处理器，存储器中存储有计算机可读指令，该处理器执行计算机可读指令时实现上述各方法实施例中的步骤。
在一些实施例中，提供了一种计算机可读存储介质，存储有计算机可读指令，该计算机可读指令被处理器执行时实现上述各方法实施例中的步骤。
在一些实施例中，还提供了一种乳腺钼靶图像的辅助诊断系统，该系统包括：乳腺DR设备、计算机设备和医生设备；乳腺DR设备与计算机设备相连，计算机设备与医生设备相连。
可选地，本申请还提供了一种包含指令的计算机程序产品，当其在计算机设备上运行时，使得计算机设备执行上述各个方法实施例所提供的乳腺钼靶图像的辅助诊断方法。
本领域普通技术人员可以理解实现上述实施例的全部或部分步骤可以通过硬件来完成，也可以通过程序来指令相关的硬件完成，所述的程序可以存储于一种计算机可读存储介质中，上述提到的存储介质可以是只读存储器，磁盘或光盘等。
以上所述仅为本申请的较佳实施例，并不用以限制本申请，凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的保护范围之内。

Claims (28)

  1. 一种乳腺钼靶图像的辅助诊断方法，由计算机设备执行，所述方法包括：
    获取单侧乳房的乳腺钼靶图像，所述乳腺钼靶图像包括：头尾CC位钼靶图像和内侧斜MLO位钼靶图像；
    调用乳房良恶性检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行良恶性预测，得到所述单侧乳房的良恶性预测结果；及
    生成和输出检测报告，所述检测报告包括所述单侧乳房的良恶性预测结果。
  2. 根据权利要求1所述的方法，其特征在于，所述乳房良恶性检测模型包括：第一单图检测部、第二单图检测部、池化层和全连接层；
    所述调用乳房良恶性检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行良恶性预测，得到所述单侧乳房的良恶性预测结果，包括：
    调用所述第一单图检测部对所述CC位钼靶图像进行处理，得到第一特征；
    调用所述第二单图检测部对所述MLO位钼靶图像进行处理，得到第二特征；及
    将所述第一特征和所述第二特征输入所述池化层和所述全连接层，得到所述单侧乳房的良恶性预测结果。
  3. 根据权利要求1所述的方法，其特征在于，所述调用乳房良恶性检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行良恶性预测，得到所述单侧乳房的良恶性预测结果之前，所述方法还包括：
    调用腺体类型分类模型对所述乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果；
    所述调用乳房良恶性检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行良恶性预测，得到所述单侧乳房的良恶性预测结果，包括：
    根据所述腺体类型识别结果确定所述乳房良恶性检测模型对应的预测阈 值；及
    调用确定所述预测阈值后的乳房良恶性检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行良恶性预测，得到所述单侧乳房的良恶性预测结果。
  4. 根据权利要求3所述的方法，其特征在于，所述调用腺体类型分类模型对所述乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果，包括：
    调用所述腺体类型分类模型对所述CC位钼靶图像的腺体类型进行识别，得到第一腺体类型；
    调用所述腺体类型分类模型对所述MLO位钼靶图像的腺体类型进行识别，得到第二腺体类型；及
    将所述第一腺体类型和所述第二腺体类型中腺体密度较大的一种，确定为所述单侧乳房的腺体类型。
  5. 根据权利要求1至4任一所述的方法，其特征在于，所述方法还包括：
    调用病灶识别模型对所述乳腺钼靶图像进行病灶检测，得到病灶检测结果，所述病灶检测包括肿块检测和钙化检测中的至少一种。
  6. 根据权利要求5所述的方法，其特征在于，所述病灶识别模型包括：病灶描述模型、病灶良恶性模型、病灶匹配模型和病灶象限定位模型中的至少一种；所述调用病灶识别模型对所述乳腺钼靶图像进行病灶检测，得到病灶检测结果，包括以下步骤中的至少一种：
    调用所述病灶描述模型对所述乳腺钼靶图像中的病灶进行检测，得到病灶描述信息；
    调用所述病灶良恶性模型对所述乳腺钼靶图像中的病灶进行良恶性识别，得到病灶良恶性概率；
    调用所述病灶匹配模型对所述CC位钼靶图像和所述MLO位钼靶图像的病灶进行一致性判断，得到病灶匹配概率；及
    调用所述病灶象限定位模型对所述乳腺钼靶图像中的病灶进行象限计 算。
  7. 根据权利要求6所述的方法，其特征在于，所述调用所述病灶象限定位模型对所述乳腺钼靶图像中的病灶进行象限计算之前，所述方法还包括：
    调用乳头检测模型对所述乳腺钼靶图像中的乳头位置进行识别；
    所述调用所述病灶象限定位模型对所述乳腺钼靶图像中的病灶进行象限计算，包括：
    调用所述病灶象限定位模型根据所述乳头位置对所述乳腺钼靶图像中的病灶进行象限计算。
  8. 根据权利要求1至4任一所述的方法，其特征在于，所述方法还包括：
    当所述乳腺钼靶图像是MLO位钼靶图像时，调用肌肉检测模型对所述乳腺钼靶图像中的肌肉位置进行识别。
  9. 根据权利要求1至4任一所述的方法，其特征在于，所述方法还包括：
    接收报告校正请求；及
    根据所述报告校正请求对所述检测报告进行校正。
  10. 根据权利要求1至4任一所述的方法，其特征在于，所述方法还包括：
    接收局部查询请求，所述局部查询请求用于请求查询所述乳腺钼靶图像中的局部区域；及
    根据所述局部查询请求输出所述局部区域对应的检测报告。
  11. 一种乳腺钼靶图像处理方法，由计算机设备执行，所述方法包括：
    获取单侧乳房的乳腺钼靶图像，所述钼靶图像包括：头尾CC位钼靶图像和内侧斜MLO位钼靶图像；
    调用目标检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果；及
    根据所述图像检测结果生成检测报告并输出。
  12. 根据权利要求11所述的方法，其特征在于，所述目标检测模型包括： 第一单图检测部、第二单图检测部、池化层和全连接层；
    所述调用目标检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果，包括：
    调用所述第一单图检测部对所述CC位钼靶图像进行处理，得到第一特征；
    调用所述第二单图检测部对所述MLO位钼靶图像进行处理，得到第二特征；及
    将所述第一特征和所述第二特征输入所述池化层和所述全连接层，得到所述单侧乳房的图像检测结果。
  13. 根据权利要求11所述的方法，其特征在于，所述调用目标检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果之前，所述方法还包括：
    调用腺体类型分类模型对所述乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果；
    所述调用目标检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果，包括：
    根据所述腺体类型识别结果确定所述目标检测模型对应的预测阈值；及
    调用确定所述预测阈值后的目标检测模型对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果。
  14. 根据权利要求13所述的方法，其特征在于，所述调用腺体类型分类模型对所述乳腺钼靶图像中的腺体类型进行识别，得到腺体类型识别结果，包括：
    调用所述腺体类型分类模型对所述CC位钼靶图像的腺体类型进行识别，得到第一腺体类型；
    调用所述腺体类型分类模型对所述MLO位钼靶图像的腺体类型进行识别，得到第二腺体类型；及
    将所述第一腺体类型和所述第二腺体类型中腺体密度较大的一种，确定 为所述单侧乳房的腺体类型。
  15. 根据权利要求11至14任一所述的方法，其特征在于，所述方法还包括：
    调用异常识别模型对所述乳腺钼靶图像进行异常检测，得到异常检测结果，所述异常检测包括肿块检测和钙化检测中的至少一种。
  16. 根据权利要求15所述的方法，其特征在于，所述异常识别模型包括：异常描述模型、异常分类模型、异常匹配模型和异常象限定位模型中的至少一种；所述调用异常识别模型对所述乳腺钼靶图像进行异常检测，得到异常检测结果，包括以下步骤中的至少一种：
    调用所述异常描述模型对所述乳腺钼靶图像中的异常区域进行检测，得到异常描述信息；
    调用所述异常分类模型对所述乳腺钼靶图像中的异常类别进行识别，得到对应的类别概率；
    调用所述异常匹配模型对所述CC位钼靶图像和所述MLO位钼靶图像的异常区域进行一致性判断，得到异常区域匹配概率；及
    调用所述异常象限定位模型对所述乳腺钼靶图像中的异常区域进行象限计算。
  17. 根据权利要求16所述的方法，其特征在于，所述调用所述异常象限定位模型对所述乳腺钼靶图像中的异常区域进行象限计算之前，所述方法还包括：
    调用乳头检测模型对所述乳腺钼靶图像中的乳头位置进行识别；
    所述调用所述异常象限定位模型对所述乳腺钼靶图像中的异常区域进行象限计算，包括：
    调用所述异常象限定位模型根据所述乳头位置对所述乳腺钼靶图像中的异常区域进行象限计算。
  18. 根据权利要求11至14任一所述的方法，其特征在于，所述方法还包括：
    当所述乳腺钼靶图像是MLO位钼靶图像时，调用肌肉检测模型对所述乳腺钼靶图像中的肌肉位置进行识别。
  19. 根据权利要求11至14任一所述的方法，其特征在于，所述方法还包括：
    接收报告校正请求；及
    根据所述报告校正请求对所述检测报告进行校正。
  20. 根据权利要求11至14任一所述的方法，其特征在于，所述方法还包括：
    接收局部查询请求，所述局部查询请求用于请求查询所述乳腺钼靶图像中的局部区域；及
    根据所述局部查询请求输出所述局部区域对应的检测报告。
  21. 一种乳腺钼靶图像的辅助诊断装置，其特征在于，所述装置包括：
    图像获取模块，用于获取单侧乳房的乳腺钼靶图像，所述乳腺钼靶图像包括：头尾CC位钼靶图像和内侧斜MLO位钼靶图像；
    乳房良恶性检测模型，用于对所述CC位钼靶图像和所述MLO位钼靶图像进行良恶性预测，得到所述单侧乳房的良恶性预测结果；及
    自动化报告输出模块，用于生成和输出检测报告，所述检测报告包括所述单侧乳房的良恶性预测结果。
  22. 一种乳腺钼靶图像处理装置，其特征在于，所述装置包括：
    图像获取模块，用于获取单侧乳房的乳腺钼靶图像，所述乳腺钼靶图像包括：头尾CC位钼靶图像和内侧斜MLO位钼靶图像；
    目标检测模型，用于对所述CC位钼靶图像和所述MLO位钼靶图像进行处理，得到所述单侧乳房的图像检测结果；及
    自动化报告输出模块，用于根据所述图像检测结果生成检测报告并输出。
  23. 一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现权利要求1至10任一所述的乳腺钼靶图像的辅助诊断方法的步骤。
  24. 一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现权利要求11至20任一所述的乳腺钼靶图像处理方法的步骤。
  25. 一种乳腺钼靶图像的辅助诊断系统，其特征在于，所述系统包括：乳腺DR设备、计算机设备和医生设备；所述乳腺DR设备与所述计算机设备相连，所述计算机设备与所述医生设备相连；所述计算机设备是如权利要求23所述的计算机设备。
  26. 一种乳腺钼靶图像处理系统，其特征在于，所述系统包括：乳腺DR设备、计算机设备和医生设备；所述乳腺DR设备与所述计算机设备相连，所述计算机设备与所述医生设备相连；所述计算机设备是如权利要求24所述的计算机设备。
  27. 一种计算机可读存储介质，存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现权利要求1至10任一所述的乳腺钼靶图像的辅助诊断方法的步骤。
  28. 一种计算机可读存储介质，存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现权利要求11至20任一所述的乳腺钼靶图像处理方法的步骤。
PCT/CN2020/088759 2019-05-16 2020-05-06 乳腺钼靶图像处理方法、装置、系统及介质 Ceased WO2020228570A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/367,266 US11922654B2 (en) 2019-05-16 2021-07-02 Mammographic image processing method and apparatus, system and medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910407807.XA CN110136829A (zh) 2019-05-16 2019-05-16 乳腺钼靶图像的辅助诊断方法、装置、系统及介质
CN201910407807.X 2019-05-16

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/367,266 Continuation US11922654B2 (en) 2019-05-16 2021-07-02 Mammographic image processing method and apparatus, system and medium

Publications (1)

Publication Number Publication Date
WO2020228570A1 true WO2020228570A1 (zh) 2020-11-19

Family

ID=67574623

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/088759 Ceased WO2020228570A1 (zh) 2019-05-16 2020-05-06 乳腺钼靶图像处理方法、装置、系统及介质

Country Status (3)

Country Link
US (1) US11922654B2 (zh)
CN (2) CN110459319B (zh)
WO (1) WO2020228570A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP4273881A4 (en) * 2021-02-09 2024-07-24 Lunit Inc. METHOD AND SYSTEM FOR PREDICTING THE RISK OF OCCURRENCE OF LESIONS

Families Citing this family (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110459319B (zh) * 2019-05-16 2021-05-25 腾讯科技（深圳）有限公司 基于人工智能的乳腺钼靶图像的辅助诊断系统
CN110689038B (zh) * 2019-06-25 2024-02-02 深圳市腾讯计算机系统有限公司 神经网络模型的训练方法、装置和医学图像处理系统
CN110738655B (zh) * 2019-10-23 2024-04-26 腾讯科技（深圳）有限公司 影像报告生成方法、装置、终端及存储介质
CN111353975A (zh) * 2020-02-24 2020-06-30 北京推想科技有限公司 网络模型训练方法及装置、病灶定位方法及装置
CN111489318B (zh) * 2020-03-20 2023-10-27 上海联影智能医疗科技有限公司 医学图像增强方法和计算机可读存储介质
CN111524106B (zh) * 2020-04-13 2021-05-28 推想医疗科技股份有限公司 颅骨骨折检测和模型训练方法、装置、设备和存储介质
US20220058794A1 (en) * 2020-08-19 2022-02-24 International Business Machines Corporation Performing diagnostic assessments
CN111709950B (zh) * 2020-08-20 2020-11-06 成都金盘电子科大多媒体技术有限公司 一种乳腺钼靶ai辅助筛查方法
US11710231B2 (en) * 2020-08-31 2023-07-25 Ping An Technology (Shenzhen) Co., Ltd. Method and apparatus for mammographic multi-view mass identification
CN112233198B (zh) * 2020-09-02 2023-04-07 山东师范大学 一种乳腺钼靶图像到能谱重组图像合成方法及系统
CN112184743B (zh) * 2020-10-29 2024-07-23 中国科学院计算技术研究所厦门数据智能研究院 一种乳腺钼靶图像中胸肌和乳头区域的分割预标注方法
CN113662573B (zh) * 2021-09-10 2023-06-30 上海联影医疗科技股份有限公司 乳腺病灶定位方法、装置、计算机设备和存储介质
JP7703408B2 (ja) * 2021-09-15 2025-07-07 富士フイルム株式会社 画像処理装置、学習装置、画像処理方法、学習方法、画像処理プログラム、及び学習プログラム
CN114068013B (zh) * 2021-11-16 2022-09-23 高峰 一种脑动脉闭塞人工智能辅助决策系统
CN114862842A (zh) * 2022-06-06 2022-08-05 北京医准智能科技有限公司 图像处理装置、电子设备及介质
CN114820590B (zh) * 2022-06-06 2023-04-07 北京医准智能科技有限公司 图像处理方法、装置、电子设备及介质
CN115170505B (zh) * 2022-07-04 2025-12-09 山东建筑大学 基于多视角自监督深度学习乳腺钼靶图像分割方法及系统
CN115359005B (zh) * 2022-08-24 2025-07-25 北京联影智能影像技术研究院 图像预测模型生成方法、装置、计算机设备和存储介质
CN115511844B (zh) * 2022-09-30 2025-09-23 上海联影智能医疗科技有限公司 病灶位置确定方法及装置
AU2024246843A1 (en) * 2023-03-31 2025-10-16 Plainsight Technologies Inc. Systems and methods for image labeling utilizing multi-model large language models
CN116977301B (zh) * 2023-07-21 2025-05-06 河南大学第一附属医院 一种基于乳腺钼靶的乳房填充物预测方法及系统
CN119206334B (zh) * 2024-09-10 2025-09-16 中电信人工智能科技(北京)有限公司 图像分类方法及装置、计算机程序产品、电子设备
CN121215274A (zh) * 2025-11-26 2025-12-26 江西乳康科技有限公司 一种乳腺癌风险预测系统

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070003119A1 (en) * 2005-07-01 2007-01-04 R2 Technology, Inc. Displaying and navigating computer-aided detection results on a review workstation
CN106203488A (zh) * 2016-07-01 2016-12-07 福州大学 一种基于受限玻尔兹曼机的乳腺图像特征融合方法
CN108392215A (zh) * 2018-02-09 2018-08-14 南方医科大学 一种同侧异位乳腺钼靶图像的点位置关联方法
CN108765387A (zh) * 2018-05-17 2018-11-06 杭州电子科技大学 基于Faster RCNN乳腺DBT影像肿块自动检测方法
CN109598702A (zh) * 2018-10-30 2019-04-09 南方医科大学南方医院 对比增强能谱乳腺x线摄影的病灶特征量化方法及系统
CN109993170A (zh) * 2019-05-10 2019-07-09 图兮深维医疗科技（苏州）有限公司 一种乳腺病灶钟形图展示装置及设备
CN110136829A (zh) * 2019-05-16 2019-08-16 腾讯科技（深圳）有限公司 乳腺钼靶图像的辅助诊断方法、装置、系统及介质

Family Cites Families (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006051198A (ja) * 2004-08-12 2006-02-23 Fuji Photo Film Co Ltd 医用画像処理システム
JP2009078033A (ja) * 2007-09-27 2009-04-16 Fujifilm Corp 乳房画像表示装置およびそのプログラム
CN101783026B (zh) * 2010-02-03 2011-12-07 北京航空航天大学 三维人脸肌肉模型的自动构造方法
KR20120072961A (ko) * 2010-12-24 2012-07-04 삼성전자주식회사 의료 영상을 이용한 영상진단을 보조하는 방법 및 장치, 이를 수행하는 영상진단 시스템
US8759783B1 (en) * 2012-12-27 2014-06-24 Ge Medical Systems Israel, Ltd. Apparatus and method for reducing examination time in molecular breast imaging
KR20140087213A (ko) * 2012-12-28 2014-07-09 삼성전자주식회사 엑스선 촬영 장치 및 입체 영상 생성 방법
EP3014572B1 (en) * 2013-06-28 2019-10-16 Koninklijke Philips N.V. Linking breast lesion locations across imaging studies
WO2015118033A1 (en) * 2014-02-04 2015-08-13 Koninklijke Philips N.V. Generating a breast parameter map
JP6692834B2 (ja) * 2015-02-09 2020-05-13 スリングショット バイオサイエンシーズ， インコーポレイテッド 調整可能な光学特性を有するヒドロゲル粒子およびその使用の方法
WO2016177691A1 (en) * 2015-05-07 2016-11-10 Agfa Healthcare A workflow management system and related method for multi-author reporting
CN105125171A (zh) * 2015-06-23 2015-12-09 牛兆河 一种用于诊断乳腺癌的系统
CN107103605B (zh) * 2016-02-22 2021-03-16 上海联影医疗科技股份有限公司 一种乳房组织的分割方法
CN105957092B (zh) * 2016-05-31 2018-10-30 福州大学 用于计算机辅助诊断的乳腺钼靶图像特征自学习提取方法
CN106339591B (zh) * 2016-08-25 2019-04-02 汤一平 一种基于深度卷积神经网络的预防乳腺癌自助健康云服务系统
US10463316B2 (en) * 2016-09-12 2019-11-05 Canon Medical Systems Corporation Medical information processing system
JP2020533039A (ja) * 2017-08-07 2020-11-19 イマーゴ・システムズ，インコーポレーテッド 画像中の物体の視覚化および特徴づけのためのシステムおよび方法
CN107798679B (zh) * 2017-12-11 2021-04-27 福建师范大学 乳腺钼靶图像乳房区域分割与钙化点检测方法
CN108416360B (zh) * 2018-01-16 2022-03-29 华南理工大学 基于乳腺钼靶钙化特征的癌症诊断系统及方法
SI3806744T1 (sl) * 2018-06-14 2022-07-29 Kheiron Medical Technologies Ltd Takojšnja ocena
JP6910323B2 (ja) * 2018-06-26 2021-07-28 富士フイルム株式会社 画像処理装置、画像処理方法、及び画像処理プログラム
EP3830837A1 (en) * 2018-08-02 2021-06-09 Cyrcadia Data Services (CDS) Limited Systems and methods for tissue assessment
US11011257B2 (en) * 2018-11-21 2021-05-18 Enlitic, Inc. Multi-label heat map display system
CN109493343A (zh) * 2018-12-29 2019-03-19 上海鹰瞳医疗科技有限公司 医疗图像异常区域分割方法及设备
US11653900B2 (en) * 2019-04-04 2023-05-23 Koninklijke Philips N.V. Data augmentation for training deep learning models with ultrasound images
WO2021183807A1 (en) * 2020-03-11 2021-09-16 The Broad Institute, Inc. Novel cas enzymes and methods of profiling specificity and activity

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070003119A1 (en) * 2005-07-01 2007-01-04 R2 Technology, Inc. Displaying and navigating computer-aided detection results on a review workstation
CN106203488A (zh) * 2016-07-01 2016-12-07 福州大学 一种基于受限玻尔兹曼机的乳腺图像特征融合方法
CN108392215A (zh) * 2018-02-09 2018-08-14 南方医科大学 一种同侧异位乳腺钼靶图像的点位置关联方法
CN108765387A (zh) * 2018-05-17 2018-11-06 杭州电子科技大学 基于Faster RCNN乳腺DBT影像肿块自动检测方法
CN109598702A (zh) * 2018-10-30 2019-04-09 南方医科大学南方医院 对比增强能谱乳腺x线摄影的病灶特征量化方法及系统
CN109993170A (zh) * 2019-05-10 2019-07-09 图兮深维医疗科技（苏州）有限公司 一种乳腺病灶钟形图展示装置及设备
CN110136829A (zh) * 2019-05-16 2019-08-16 腾讯科技（深圳）有限公司 乳腺钼靶图像的辅助诊断方法、装置、系统及介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZHU JIANPING (TENCENT TECHNICAL ENGINEERING OFFICIAL ACCOUNT): "Tencent's exploration in the direction of breast cancer imaging AI diagnosis", COLUMN HOME - COLUMN OF TENCENT TECHNICAL ENGINEERING OFFICIAL ACCOUNT - TENCENT'S EXPLORATION IN THE DIRECTION OF BREAST CANCER IMAGING AI DIAGNOSIS, 23 April 2018 (2018-04-23), XP009524278 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP4273881A4 (en) * 2021-02-09 2024-07-24 Lunit Inc. METHOD AND SYSTEM FOR PREDICTING THE RISK OF OCCURRENCE OF LESIONS

Also Published As

Publication number Publication date
US20210338179A1 (en) 2021-11-04
CN110459319B (zh) 2021-05-25
CN110459319A (zh) 2019-11-15
CN110136829A (zh) 2019-08-16
US11922654B2 (en) 2024-03-05

Similar Documents

Publication Publication Date Title
WO2020228570A1 (zh) 乳腺钼靶图像处理方法、装置、系统及介质
US11880972B2 (en) Tissue nodule detection and tissue nodule detection model training method, apparatus, device, and system
CN111325739B (zh) 肺部病灶检测的方法及装置，和图像检测模型的训练方法
CN109447065B (zh) 一种乳腺影像识别的方法及装置
EP3796210A1 (en) Spatial distribution of pathological image patterns in 3d image data
WO2021179491A1 (zh) 图像处理方法、装置、计算机设备和存储介质
WO2021021329A1 (en) System and method for interpretation of multiple medical images using deep learning
Thakur et al. A systematic review of machine and deep learning techniques for the identification and classification of breast cancer through medical image modalities
US9600628B2 (en) Automatic generation of semantic description of visual findings in medical images
CN110009600A (zh) 一种医学图像区域过滤方法、装置及存储介质
WO2021128825A1 (zh) 三维目标检测及模型的训练方法及装置、设备、存储介质
CN110046627B (zh) 一种乳腺影像识别的方法及装置
CN108464840A (zh) 一种乳腺肿块自动检测方法及系统
US20160203600A1 (en) Methods and systems for determining breast density
US10706534B2 (en) Method and apparatus for classifying a data point in imaging data
JP7346553B2 (ja) 深層学習を使用する３ｄデータセット内のオブジェクトの成長率の決定
Hu et al. A multi-instance networks with multiple views for classification of mammograms
WO2022033598A1 (zh) 乳房x射线图像获取方法、装置、计算机设备和存储介质
WO2022138277A1 (ja) 学習装置、方法及びプログラム並びに医用画像処理装置
CN111489318A (zh) 医学图像增强方法和计算机可读存储介质
CN114287043A (zh) 使用减少的参考图像分析和减少站点间/站点内变异性以及改进放射科医生诊断准确性和一致性的方法
CN109461144B (zh) 一种乳腺影像识别的方法及装置
Aishwarya et al. A survey on comparative study of lung nodules applying machine learning and deep learning techniques
CN114550882B (zh) 改善磁共振成像注释
CN115661176A (zh) 基于ct图像的三维体构建方法、装置、存储介质以及终端

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20805693

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20805693

Country of ref document: EP

Kind code of ref document: A1