WO2019223147A1 - 肝脏癌变定位方法、装置及存储介质 - Google Patents
肝脏癌变定位方法、装置及存储介质 Download PDFInfo
- Publication number
- WO2019223147A1 WO2019223147A1 PCT/CN2018/102133 CN2018102133W WO2019223147A1 WO 2019223147 A1 WO2019223147 A1 WO 2019223147A1 CN 2018102133 W CN2018102133 W CN 2018102133W WO 2019223147 A1 WO2019223147 A1 WO 2019223147A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- preset
- slice
- generate
- images
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/24—Aligning, centring, orientation detection or correction of the image
- G06V10/245—Aligning, centring, orientation detection or correction of the image by locating a pattern; Special marks for positioning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
- G06V2201/031—Recognition of patterns in medical or anatomical images of internal organs
Definitions
- the present application relates to the field of picture recognition technology, and in particular, to a method, a device, and a computer-readable storage medium for locating liver cancer.
- CT Computer Tomography
- traditional methods rely on the experience of doctors to judge multiple CT pictures, locate the location of the lesion, and the speed and accuracy of liver cancer localization are greatly affected by the experience of doctors.
- the CT image is a grayscale image and the same CT image shows multiple internal organs, at the same time, the number of CT slice images related to the liver is large, resulting in a doctor consuming extreme brain power and inefficient lesion localization. Therefore, how to quickly and accurately locate the location of liver canceration has become an urgent technical problem.
- the present application provides a liver cancer localization method, device, and computer-readable storage medium, the main purpose of which is to use artificial intelligence detection technology to perform rapid localization detection of liver cancerous positions on CT slice images, and improve liver cancerous localization speed .
- liver cancer localization method which includes:
- Sample processing step Obtain a first preset number of CT slice sample images. Each CT slice sample image is marked with a lesion marker point and a lesion shape curve defined by the lesion marker point. Each CT slice sample image is marked with a non- Cancer markers or cancer markers, and pre-process each acquired CT slice sample image to generate a corresponding pre-processed image;
- Deformation step Generate a corresponding deformation image for each pre-processed image according to a preset deformation rule, and form each corresponding pre-processed image and its corresponding deformation image into a corresponding set of images to be trained;
- Training steps use the images in the image collection to train the recognition model
- Receiving step receiving a CT slice image of a liver cancer position location
- the CT slice image is input into a trained recognition model to locate and identify the location of liver cancer.
- the present application also provides an electronic device including a memory and a processor.
- the memory stores a liver cancer localization program, and the liver cancer localization program is executed by the processor, which can implement the following steps:
- Sample processing step Obtain a first preset number of CT slice sample images. Each CT slice sample image is marked with a lesion marker point and a lesion shape curve defined by the lesion marker point. Each CT slice sample image is marked with a non- Cancer markers or cancer markers, and pre-process each acquired CT slice sample image to generate a corresponding pre-processed image;
- Deformation step Generate corresponding deformation images for each pre-processed image according to preset deformation rules, and form each corresponding pre-processed image and its corresponding deformation image into a corresponding set of images to be trained;
- Training steps use the images in the image collection to train the recognition model
- Receiving step receiving a CT slice image of a liver cancer position location
- the CT slice image is input into a trained recognition model to locate and identify the location of liver cancer.
- the present application also provides a computer-readable storage medium, where the computer-readable storage medium includes a liver cancer localization program, and when the liver cancer localization program is executed by a processor, the foregoing can be achieved. Arbitrary steps in liver cancer localization methods.
- the liver canceration localization method, electronic device, and computer-readable storage medium provided in the present application receive the CT slice image to be localized for liver cancerization, use a pre-trained recognition model to locate the liver cancer change position on the CT slice image, and Placing a cancerous label on a place with cancerousness, thereby improving the accuracy of localizing liver cancerousness on CT slice images, reducing labor costs, and improving work efficiency.
- FIG. 1 is a schematic diagram of a preferred embodiment of an electronic device of the present application
- FIG. 2 is a schematic block diagram of a preferred embodiment of a liver cancer localization procedure in FIG. 1;
- FIG. 3 is a flowchart of a preferred embodiment of a liver cancer localization method of the present application.
- FIG. 4 is a flowchart of training of an identification model of the present application.
- FIG. 1 it is a schematic diagram of a preferred embodiment of the electronic device 1 of the present application.
- the electronic device 1 may be a server, a smart phone, a tablet computer, a personal computer, a portable computer, and other electronic devices with computing functions.
- the electronic device 1 includes a memory 11, a processor 12, a network interface 13 and a communication bus 14.
- the network interface 13 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).
- the communication bus 14 is used to implement connection communication between these components.
- the memory 11 includes at least one type of readable storage medium.
- the at least one type of readable storage medium may be a non-volatile storage medium such as a flash memory, a hard disk, a multimedia card, a card-type memory, and the like.
- the memory 11 may be an internal storage unit of the electronic device 1, such as a hard disk of the electronic device 1.
- the memory 11 may also be an external storage unit of the electronic device 1, such as a plug-in hard disk, a Smart Memory Card (SMC) provided on the electronic device 1, and security. Digital (Secure Digital, SD) card, Flash card (Flash card), etc.
- the memory 11 may not only be used to store application software installed on the electronic device 1 and various types of data, such as a liver cancer localization program 10, a CT slice image to be localized for cancer, and an electronic computer for model training. Computed Tomography (CT) slice sample images.
- CT Computed Tomography
- the processor 12 may be a central processing unit (CPU), a microprocessor or other data processing chip in some embodiments, and is configured to run program code or process data stored in the memory 11, for example, to perform liver cancer localization. Computer program code of the program 10, training of the recognition model, and the like.
- CPU central processing unit
- microprocessor or other data processing chip in some embodiments, and is configured to run program code or process data stored in the memory 11, for example, to perform liver cancer localization.
- Computer program code of the program 10 training of the recognition model, and the like.
- the electronic device 1 may further include a display, and the display may be referred to as a display screen or a display unit.
- the display may be an LED display, a liquid crystal display, a touch-type liquid crystal display, an organic light-emitting diode (OLED) touch device, or the like.
- the display is used to display information processed in the electronic device 1 and to display a visualized work interface.
- the electronic device 1 may further include a user interface.
- the user interface may include an input unit such as a keyboard, and a voice output device such as a stereo or a headset.
- the user interface may further include a standard wired interface and a wireless interface.
- FIG. 2 it is a schematic block diagram of a preferred embodiment of a liver cancer localization procedure in FIG. 1.
- the module referred to in this application refers to a series of computer program instruction segments capable of performing specific functions.
- the liver cancer localization program 10 includes a sample processing module 110, a deformation module 120, a training module 130, a receiving module 140, and a recognition module 150.
- the functions or operation steps implemented by the modules 110-150 are described below. :
- the sample processing module 110 is configured to obtain a first preset number of CT slice sample images.
- Each CT slice sample image is marked with a lesion mark point and a lesion shape curve defined by the lesion mark point, and each CT slice sample image corresponds to It is marked with a non-cancer marker or a cancer marker, and each acquired CT slice sample image is pre-processed to generate a corresponding pre-processed image.
- the pre-processing specifically includes: performing pixel filtering of the preset gray scale range on each CT slice sample image according to a predetermined gray scale range of the liver tissue on the CT slice image to generate a corresponding filtered image, and Ensure that the image size of each filtered image is consistent with the image size of the corresponding CT slice sample image.
- the histogram equalization processing is performed on each filtered image to generate an image after the equalization processing, and each image after the equalization processing is a preprocessed image. Further, the contrast can be enhanced according to methods such as histogram stretching.
- the deformation module 120 is configured to generate a corresponding deformation image for each pre-processed image according to a preset deformation rule, and form each corresponding pre-processed image and its corresponding deformation image into a corresponding set of images to be trained.
- the preset deformation rule includes: a pre-processed image to be deformed, adding the pre-processed image to increase Gaussian noise, and generating a corresponding noise-added image.
- the Gaussian noise is completely determined by the covariance function of its time-varying average and two instantaneous averages. Then, within the preset angle range, angularly rotate the noise-added image to generate a corresponding rotated image.
- the rotation image is elastically transformed to generate a corresponding deformation image, and each pre-processed image and its corresponding deformation image are respectively composed into a corresponding set of images to be trained.
- the preset elastic transformation rule includes: for a rotated image, each pixel point (xi, yi) on the rotated image is generated with 2 random numbers ⁇ x in a range of [-1 to 1]. (xi, yi) and ⁇ y (xi, yi), store the random number ⁇ x (xi, yi) on the xi of the pixels (xi, yi) of the pixel matrix D and E of the same size as the rotated image, representing the pixel points (xi, yi) in the x-direction moving distance, and store the random number ⁇ y (xi, yi) on the yi of the pixels (xi, yi) of the pixel matrix D and E of the same size as the rotated image, which represents the pixel point
- the moving distance in the y direction of (xi, yi) yields two random number matrices D1 and E1.
- the range includes, but is not limited to, [-1 to 1].
- the two convolution result images are applied to the original image: the pixels at the position (xi, yi) of the rotated image are placed in the new image (xi + A (xi, yi), yi + B (xi, yi)) Position to move all pixels to get the final deformed image.
- the training module 130 is configured to train the recognition model by using the images in the image set.
- the pre-trained recognition model is a Convolutional Neural Network (CNN) model.
- CNN Convolutional Neural Network
- the convolutional neural network model superimposes features of the same dimension in the convolution step in the upsampling step, and then passes through the compressed space.
- the convolution operation compresses the image to obtain an image with the same feature space as before the superposition.
- the model structure of the convolutional neural network model is shown in Table 1.
- the operating principle of the convolutional neural network model of the preset structure is as follows:
- Each sample input is a 512 * 512 * n pre-processed image, where n is the number of CT slices of the sample.
- the maximum value is pooled, using 2 * 2 cores, and outputting 256 * 256 size;
- the maximum value is pooled, using 2 * 2 cores, and outputting 128 * 128 size;
- the maximum value is pooled, using 2 * 2 cores, and outputting 64 * 64 size;
- the maximum value is pooled, using 2 * 2 cores, and outputting 32 * 32 size;
- Stitching, splicing drop5 and up7, output 2048 feature maps, 32 * 32 size;
- Upsampling using 2 * 2 upsampling, output 256 * 256;
- Stitching, conv1 and up11, output 128 feature maps, 512 * 512 size
- Convolution using a 3 * 3 convolution kernel, outputs a feature map, using a sigmoid activation function, and outputs a size of 512 * 512.
- the receiving module 140 is configured to receive a CT slice image of a liver cancer position location. After receiving the CT slice image, in order to enhance the contrast and highlight the liver tissue, the received CT slice image is subjected to pixel filtering according to a predetermined gray range of the liver tissue on the CT slice image to generate a filtered image. Ensure that the image size of the filtered image is consistent with the received CT slice image size. Then, a histogram equalization process is performed on the filtered image to generate an image after the equalization process. Finally, the equalized image is input into a recognition model for localization and recognition.
- the recognition module 150 is configured to use a pre-trained recognition model to locate and identify a liver cancerous position of the CT slice image. If a liver cancerous position is identified, a label of a preset form is labeled on the liver cancerous position of the CT slice image. For example, if a certain location is identified as having liver cancer on a CT slice image, a curved wire frame is generated at the identified location of the liver cancer lesion area, and labeled in the wire frame.
- FIG. 3 it is a flowchart of a preferred embodiment of a liver cancer localization method of the present application.
- the method for realizing liver cancerous localization includes: step S10-step S50:
- the sample processing module 110 obtains a first preset number of CT slice sample images.
- Each CT slice sample image is marked with a lesion mark point and a lesion shape curve defined by the lesion mark point, and each CT slice sample image corresponds to Marked non-cancer or cancer.
- 10,000 CT slice sample images are acquired, of which 8000 CT slice sample images have liver cancer lesion areas, and 2000 CT slice sample images have no liver cancer lesion areas.
- the lesion mark point refers to a boundary point between a lesion area and a non-lesion area.
- each of the acquired CT slice sample images is preprocessed to generate corresponding preprocessed images.
- the preprocessing specifically includes: presetting a predetermined grayscale range of the liver tissue on the CT slice image, for example, the grayscale range of the liver tissue is [-100 to 400], and presetting each CT slice sample image separately. Pixel filtering in the gray range to generate corresponding filtered images, and ensure that the image size of each filtered image is consistent with the image size of the corresponding CT slice sample image. Next, the histogram equalization processing is performed on each filtered image to generate an image after the equalization processing, and each image after the equalization processing is a preprocessed image. Further, the contrast can be enhanced according to methods such as histogram stretching.
- the deformation module 120 generates a corresponding deformation image for each pre-processed image according to a preset deformation rule, and forms each corresponding pre-processed image and its corresponding deformation image into a corresponding set of images to be trained.
- the preset deformation rule includes: a pre-processed image to be deformed, adding the pre-processed image to increase Gaussian noise, and generating a corresponding noise-added image.
- the Gaussian noise is completely determined by the covariance function of its time-varying average and two instantaneous averages.
- a random number with a Gaussian distribution is randomly generated, the random number is added to the pixel value of the preprocessed image, and the added value is compressed into an interval of [0 to 225] to obtain a corresponding noise image. Then, within the preset angle range, angularly rotate the noise-added image to generate a corresponding rotated image. It is assumed that the preset angle range is [-30 to 30], a certain angle is randomly selected within the preset angle range, and the noise-added image is rotated by the angle to generate a corresponding rotated image. Finally, according to a preset elastic transformation rule, the rotation image is elastically transformed to generate a corresponding deformed image, and each preprocessed image and its corresponding deformed image are respectively composed into a corresponding set of images to be trained.
- the preset elastic transformation rule includes: for a rotated image, each pixel point (xi, yi) on the rotated image is generated with 2 random numbers ⁇ x in a range of [-1 to 1]. (xi, yi) and ⁇ y (xi, yi), store the random number ⁇ x (xi, yi) on the xi of the pixels (xi, yi) of the pixel matrix D and E of the same size as the rotated image, representing the pixel (xi, yi) in the x-direction moving distance, and store the random number ⁇ y (xi, yi) on the yi of the pixels (xi, yi) of the pixel matrix D and E of the same size as the rotated image, which represents the pixel point
- the moving distance in the y direction of (xi, yi) yields two random number matrices D1 and E1.
- the range includes, but is not limited to, [-1 to 1].
- the two convolution result images are applied to the original image: the pixels at the position (xi, yi) of the rotated image are placed in the new image (xi + A (xi, yi), yi + B (xi, yi)) Position to move all pixels to get the final deformed image.
- step S30 the training module 130 uses the images in the image set to train the recognition model.
- FIG. 4 it is a flowchart of recognition model training of the present application.
- the training steps of the recognition model are as follows:
- All the sets of images to be trained are divided into a training set of the first scale and a validation set of the second scale. For example, all the to-be-trained image sets are randomly divided into a training set and a verification set according to a ratio of 7: 3. The training set accounts for 70% of all the to-be-trained image sets, and the remaining 30% of the to-be-trained image sets are used as the verification set to detect the model.
- Model training is performed using each image in the training set to generate the recognition model, and each image generated in the verification set is used to verify the generated recognition model. For example, 7000 image sets in the training set are used to train the model, and 3000 image sets in the validation set are used to verify to generate the optimal recognition model.
- the training is completed. If the verification pass rate is less than the preset threshold, a second preset number of CT slice sample images are added, and the increased CT slice sample images are pre-processed and deformed. After processing, the flow returns to the step of dividing the image set into a training set and a validation set. It is assumed that the preset threshold is 98%, and the image set in the verification set is substituted into the recognition model for verification. If the pass rate is greater than or equal to 98%, the recognition model is the optimal model. If the pass rate is less than 98%, 2000 CT slice sample images are added, and the added CT slice sample images are pre-processed and deformed. The flow returns to the step of dividing the image set into a training set and a verification set.
- the receiving module 140 receives a CT slice image to be positioned for liver cancer.
- the received CT slice image is subjected to pixel filtering according to a predetermined gray range of the liver tissue on the CT slice image to generate a filtered image. For example, set the gray value of the liver in the gray range [-100 ⁇ 400], and filter the CT slice image.
- the image size of the filtered image is consistent with the received CT slice image size.
- the histogram equalization processing is performed on the filtered image to expand the gray level with a large number of pixels, expand the dynamic range of image element values, and generate an equalized image to highlight the contrast between the liver and other tissues in the image .
- the filtered image is subjected to a histogram equalization process.
- the equalized image is input into a recognition model for localization and recognition.
- the recognition module 150 uses a pre-trained recognition model to locate and identify the liver cancerous position of the CT slice image.
- the labeling module 130 labels the liver cancerous position in the CT slice image with a label in a preset form. For example, if a certain position is identified as having liver cancer on a CT slice image, then the position of the liver cancer lesion area generating the curve wire frame is marked: "the patient corresponding to this CT slice image has cancerous liver here".
- the liver canceration localization method proposed in this embodiment uses a pre-trained recognition model to quickly and accurately locate the liver canceration of the received CT slice image, thereby improving the speed and accuracy of liver cancerization localization.
- an embodiment of the present application further provides a computer-readable storage medium, where the computer-readable storage medium includes a liver cancer localization program 10, and the liver cancer localization program 10 implements the following operations when executed by a processor:
- Sample processing step Obtain a first preset number of CT slice sample images. Each CT slice sample image is marked with a lesion marker point and a lesion shape curve defined by the lesion marker point. Each CT slice sample image is marked with a non- Cancer markers or cancer markers, and pre-process each acquired CT slice sample image to generate a corresponding pre-processed image;
- Deformation step Generate corresponding deformation images for each pre-processed image according to preset deformation rules, and form each corresponding pre-processed image and its corresponding deformation image into a corresponding set of images to be trained;
- Training steps use the images in the image collection to train the recognition model
- Receiving step receiving a CT slice image of a liver cancer position location
- the CT slice image is input into a trained recognition model to locate and identify the location of liver cancer.
- the training steps of the pre-trained recognition model are as follows:
- the training is completed. If the verification pass rate is less than the preset threshold, a second preset number of CT slice sample images are added, and the increased CT slice sample images are preprocessed and deformed. After processing, the flow returns to the step of dividing the image set into a training set and a validation set.
- the pre-processing includes:
- pixel filtering of the preset gray scale range is performed on each CT slice sample image to generate a corresponding filtered image and ensure the image size of each filtered image Consistent with the image size of the corresponding CT slice sample image;
- Histogram equalization processing is performed on each filtered image to generate an image after the equalization processing, and each image after the equalization processing is a preprocessed image.
- the preset deformation rule is:
- the preset elastic transformation rule is:
- two random numbers ⁇ x (xi, yi) and ⁇ y (xi, yi) are generated for each pixel (xi, yi) on the rotated image in the range [-1 ⁇ 1].
- the receiving step includes:
- the pixels of the preset gray range are used to filter the received CT slice image to generate a filtered image, and ensure that the image size of the filtered image and the CT slice images have the same image size;
- a histogram equalization process is performed on the filtered image to generate an image after the equalization process.
- the specific implementation manner of the computer-readable storage medium of the present application is substantially the same as the specific implementation manner of the liver cancer localization method described above, and details are not described herein again.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
- Apparatus For Radiation Diagnosis (AREA)
Abstract
一种肝脏癌变定位方法、装置及存储介质,该方法包括:获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线及非癌症标记或者癌症标记,并对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像(S10);分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合(S20);利用图像集合中的图像对识别模型进行训练(S30);接收待进行肝脏癌变位置定位的CT切片图像(S40);将该CT切片图像输入训练好的识别模型进行肝脏癌变位置的定位识别(S50)。该方法通过对CT切片图像的识别,提高对肝脏癌变位置的检测效率及准确性。
Description
本申请要求于2018年05月23日提交中国专利局、申请号为201810501877.7,名称为“肝脏癌变定位方法、装置及存储介质”的中国专利申请的优先权,该中国专利申请的整体内容以参考的方式结合本申请中。
本申请涉及图片识别技术领域,尤其涉及一种肝脏癌变定位方法、装置及计算机可读存储介质。
目前,对于肝脏癌变的诊断是通过CT(Computed Tomography,电子计算机断层扫描)断层图像对人体肝脏横断面图像是否病变进行判断的。然而,传统的方法是凭借医生的经验对多张CT图片进行判断,定位病变位置,肝脏癌变定位的速度及准确性受医生经验的影响较大。另一方面,由于CT图像为灰度图像而且同一个CT图像显示多个内脏器官,同时,与肝脏相关的CT切片图像数量又多,导致医生消耗极大脑力且病变定位效率低下。因此,如何对肝脏癌变位置进行快速准确定位已成为一个亟待解决的技术问题。
发明内容
鉴于以上内容,本申请提供一种肝脏癌变定位方法、装置及计算机可读存储介质,其主要目的在于利用人工智能检测技术对CT切片图像上的肝脏癌变位置进行快速定位检测,提高肝脏癌变定位速度。
为实现上述目的,本申请提供一种肝脏癌变定位方法,该方法包括:
样本处理步骤:获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线,每个CT切片样本图像对应标有非癌症标记或者癌症标记,并对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像;
形变步骤:分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练 图像集合;
训练步骤:利用图像集合中的图像对识别模型进行训练;
接收步骤:接收待进行肝脏癌变位置定位的CT切片图像;
识别步骤:将该CT切片图像输入训练好的识别模型进行肝脏癌变位置的定位识别。
此外,本申请还提供一种电子装置,该电子装置包括:存储器、处理器,所述存储器上存储肝脏癌变定位程序,所述肝脏癌变定位程序被所述处理器执行,可实现如下步骤:
样本处理步骤:获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线,每个CT切片样本图像对应标有非癌症标记或者癌症标记,并对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像;
形变步骤:分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合;
训练步骤:利用图像集合中的图像对识别模型进行训练;
接收步骤:接收待进行肝脏癌变位置定位的CT切片图像;
识别步骤:将该CT切片图像输入训练好的识别模型进行肝脏癌变位置的定位识别。
此外,为实现上述目的,本申请还提供一种计算机可读存储介质,所述计算机可读存储介质中包括肝脏癌变定位程序,所述肝脏癌变定位程序被处理器执行时,可实现如上所述肝脏癌变定位方法中的任意步骤。
本申请提出的肝脏癌变定位方法、电子装置及计算机可读存储介质,通过接收待肝脏癌变定位的CT切片图像,利用预先训练的识别模型对该CT切片图像上的肝脏癌变为位置进行定位,并在有癌变的位置贴上癌变标签,从而提高CT切片图像上肝脏癌变的定位准确度,减少人力成本,提高工作效率。
图1为本申请电子装置较佳实施例的示意图;
图2为图1中肝脏癌变定位程序较佳实施例的模块示意图;
图3为本申请肝脏癌变定位方法较佳实施例的流程图;
图4为本申请识别模型训练的流程图。
本申请目的的实现、功能特点及优点将结合实施例,参照附图做进一步说明。
应当理解,此处所描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
如图1所示,是本申请电子装置1较佳实施例的示意图。
在本实施例中,电子装置1可以是服务器、智能手机、平板电脑、个人电脑、便携计算机以及其他具有运算功能的电子设备。
该电子装置1包括:存储器11、处理器12、网络接口13及通信总线14。其中,网络接口13可选地可以包括标准的有线接口、无线接口(如WI-FI接口)。通信总线14用于实现这些组件之间的连接通信。
存储器11至少包括一种类型的可读存储介质。所述至少一种类型的可读存储介质可为如闪存、硬盘、多媒体卡、卡型存储器等的非易失性存储介质。在一些实施例中,所述存储器11可以是所述电子装置1的内部存储单元,例如该电子装置1的硬盘。在另一些实施例中,所述存储器11也可以是所述电子装置1的外部存储单元,例如所述电子装置1上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。
在本实施例中,所述存储器11不仅可以用于存储安装于所述电子装置1的应用软件及各类数据,例如肝脏癌变定位程序10、待癌变定位的CT切片图像和模型训练的电子计算机断层扫描(Computed Tomography,CT)切片样本图像。
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit,CPU),微处理器或其它数据处理芯片,用于运行存储器11中存储的程序代码或处理数据,例如执行肝脏癌变定位程序10的计算机程序代码和识别模型的训练等。
优选地,该电子装置1还可以包括显示器,显示器可以称为显示屏或显 示单元。在一些实施例中显示器可以是LED显示器、液晶显示器、触控式液晶显示器以及有机发光二极管(Organic Light-Emitting Diode,OLED)触摸器等。显示器用于显示在电子装置1中处理的信息以及用于显示可视化的工作界面。
优选地,该电子装置1还可以包括用户接口,用户接口可以包括输入单元比如键盘(Keyboard)、语音输出装置比如音响、耳机等,可选地用户接口还可以包括标准的有线接口、无线接口。
如图2所示,是图1中肝脏癌变定位程序较佳实施例的模块示意图。本申请所称的模块是指能够完成特定功能的一系列计算机程序指令段。
在本实施例中,肝脏癌变定位程序10包括:样本处理模块110、形变模块120、训练模块130、接收模块140、识别模块150,所述模块110-150所实现的功能或操作步骤如下所述:
样本处理模块110,用于获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线,每个CT切片样本图像对应标有非癌症标记或者癌症标记,并对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像。所述预处理具体包括:根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,分别对各个CT切片样本图像进行预设灰度范围的像素过滤,以生成对应的过滤图像,并保证各个过滤图像的图像尺寸与对应的CT切片样本图像的图像尺寸一致。接着,分别对各个过滤图像进行直方图均衡化处理,生成均衡化处理后的图像,各个均衡化处理后的图像即为预处理图像。进一步地,还可以根据直方图拉伸等方法增强对比度。
形变模块120,用于分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合。其中,所述预设的形变规则为包括:对待形变处理的预处理图像,增加该预处理图像增加高斯噪声,生成对应的加噪图像。所述高斯噪声完全由其时变平均值和两瞬时平均值的协方差函数确定。接着,在预设角度范围内,对该加噪图像进行角度随机旋转,生成对应的旋转图像。最后,根据预设的弹性变换规则,对该旋转图像进行弹性变换,生成对应的形变图 像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合。
其中,所述预设的弹性变换规则包括:针对一个旋转图像,分别对该旋转图像上的每个像素点(xi,yi)在范围为[-1~1]之间产生2个随机数Δx(xi,yi)和Δy(xi,yi),将随机数Δx(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的xi上,表示像素点(xi,yi)的x方向的移动距离,并将随机数Δy(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的yi上,表示像素点(xi,yi)的y方向的移动距离,得到2个随机数矩阵D1和E1。但应理解的是,所述范围包括但不限于[-1~1]。然后,随机生成一个以第一预设值为均值,以第二预设值为标准差的预设大小为105*105的高斯核,将该高斯核分别与随机数矩阵D1和E1分别卷积,生成2个卷积结果图像,分别为A(xi,yi)、B(xi,yi)。最后,将2个卷积结果图像作用于原图像:将该旋转图像的位置(xi,yi)的像素放到新图(xi+A(xi,yi),yi+B(xi,yi))位置上,以将所有像素移动之后得到最后的形变图像。
训练模块130,用于利用图像集合中的图像对识别模型进行训练。其中,所述预先训练的识别模型为卷积神经网络(Convolutional Neural Network,CNN)模型,该卷积神经网络模型在上采样步骤中叠加卷积步骤中相同维度的特征,之后再经过压缩空间的卷积操作对图像进行压缩,得到与叠加前相同特征空间的图像。该卷积神经网络模型的模型结构如表1所示。
表1:识别模型的网络结构
层标签 | 功能层名 | 卷积核或参数 | 特征图个数 | 输出大小 |
卷积 | 3*3 | 64 | 512*512*64 | |
Conv1 | 卷积 | 3*3 | 64 | 512*512*64 |
最大值池化 | 2*2 | \ | 256*256*64 | |
卷积 | 3*3 | 128 | 256*256*128 | |
Conv2 | 卷积 | 3*3 | 128 | 256*256*128 |
最大值池化 | 2*2 | \ | 128*128*128 | |
卷积 | 3*3 | 256 | 128*128*256 | |
Conv3 | 卷积 | 3*3 | 256 | 128*128*256 |
最大值池化 | 2*2 | \ | 64*64*256 | |
卷积 | 3*3 | 512 | 64*64*512 | |
卷积 | 3*3 | 512 | 64*64*512 | |
Drop4 | 丢弃 | 0.5 | \ | 64*64*512 |
最大值池化 | 2*2 | \ | 32*32*512 | |
卷积 | 3*3 | 1024 | 32*32*1024 | |
卷积 | 3*3 | 1024 | 32*32*1024 | |
Drop5 | 丢弃 | 0.5 | \ | 32*32*1024 |
最大值池化 | 2*2 | \ | 16*16*1024 | |
卷积 | 3*3 | 2048 | 16*16*2048 | |
卷积 | 3*3 | 2048 | 16*16*2048 | |
Drop6 | 丢弃 | 0.5 | \ | 16*16*2048 |
上采样 | 2*2 | \ | 32*32*2048 | |
Up7 | 卷积 | 2*2 | 1024 | 32*32*1024 |
拼接 | Drop5 | 2048 | 32*32*2048 | |
卷积 | 3*3 | 1024 | 32*32*1024 | |
卷积 | 3*3 | 1024 | 32*32*1024 | |
上采样 | 2*2 | \ | 64*64*1024 | |
Up8 | 卷积 | 2*2 | 512 | 64*64*512 |
拼接 | Drop4 | 1024 | 64*64*1024 | |
卷积 | 3*3 | 512 | 64*64*512 | |
卷积 | 3*3 | 512 | 64*64*512 | |
上采样 | 2*2 | \ | 128*128*512 | |
Up9 | 卷积 | 2*2 | 256 | 128*128*256 |
拼接 | Conv3 | 512 | 128*128*512 | |
卷积 | 3*3 | 256 | 128*128*256 | |
卷积 | 3*3 | 256 | 128*128*256 | |
上采样 | 2*2 | \ | 256*256*256 | |
Up10 | 卷积 | 2*2 | 128 | 256*256*128 |
拼接 | Conv2 | 256 | 256*256*256 | |
卷积 | 3*3 | 128 | 256*256*128 | |
卷积 | 3*3 | 128 | 256*256*128 | |
上采样 | 2*2 | \ | 512*512*128 | |
Up11 | 卷积 | 2*2 | 64 | 512*512*64 |
拼接 | Conv1 | 128 | 512*512*128 | |
卷积 | 3*3 | 64 | 512*512*64 | |
卷积 | 3*3 | 64 | 512*512*64 | |
卷积 | 3*3 | 2 | 512*512*2 | |
卷积 | 3*3 | 1 | 512*512*1 |
该预设结构的卷积神经网络模型的运行原理如下:
输入的每个样本是512*512*n的预处理后图像,其中n为该样本CT切片数量。进行以下模型:
卷积,采用3*3卷积核,输出64个特征图,采用Relu激活函数,输出512*512大小;
卷积,采用3*3卷积核,输出64个特征图,采用Relu激活函数,输出512*512大小,记为conv1;
最大值池化,采用2*2核,输出256*256大小;
卷积,采用3*3卷积核,输出128个特征图,采用Relu激活函数,输出256*256大小;
卷积,采用3*3卷积核,输出128个特征图,采用Relu激活函数,输出256*256大小,记为conv2;
最大值池化,采用2*2核,输出128*128大小;
卷积,采用3*3卷积核,输出256个特征图,采用Relu激活函数,输出128*128大小;
卷积,采用3*3卷积核,输出256个特征图,采用Relu激活函数,输出128*128大小,记为conv3;
最大值池化,采用2*2核,输出64*64大小;
卷积,采用3*3卷积核,输出512个特征图,采用Relu激活函数,输出64*64大小;
卷积,采用3*3卷积核,输出512个特征图,采用Relu激活函数,输出64*64大小,记为conv4;
丢弃,随机选conv4的一半输出设为0,输出记为drop4;
最大值池化,采用2*2核,输出32*32大小;
卷积,采用3*3卷积核,输出1024个特征图,采用Relu激活函数,输出32*32大小;
卷积,采用3*3卷积核,输出1024个特征图,采用Relu激活函数,输出32*32大小,记为conv5;
丢弃,随机选conv5的一半输出设为0,输出记为drop5;
最大值池化,采用2*2核,输出16*16大小;
卷积,采用3*3卷积核,输出2048个特征图,采用Relu激活函数,输出 16*16大小;
卷积,采用3*3卷积核,输出2048个特征图,采用Relu激活函数,输出16*16大小,记为conv6;
丢弃,随机选conv6的一半输出设为0,输出记为drop6;
上采样,采用2*2上采样,输出32*32;
卷积,采用2*2卷积核,输出1024个特征图,采用Relu激活函数,输出32*32大小,记为up7;
拼接,拼接drop5和up7,输出2048个特征图,32*32大小;
卷积,采用3*3卷积核,输出1024个特征图,采用Relu激活函数,输出32*32大小;
卷积,采用3*3卷积核,输出1024个特征图,采用Relu激活函数,输出32*32大小;
上采样,采用2*2上采样,输出64*64;
卷积,采用2*2卷积核,输出512个特征图,采用Relu激活函数,输出64*64大小,记为up8;
拼接,拼接drop4和up8,输出1024个特征图,64*64大小;
卷积,采用3*3卷积核,输出512个特征图,采用Relu激活函数,输出64*64大小;
卷积,采用3*3卷积核,输出512个特征图,采用Relu激活函数,输出64*64大小;
上采样,采用2*2上采样,输出128*128大小;
卷积,采用2*2卷积核,输出256个特征图,采用Relu激活函数,输出128*128大小,记为up9;
拼接,拼接conv3和up9,输出512个特征图,128*128大小;
卷积,采用3*3卷积核,输出256个特征图,采用Relu激活函数,输出128*128大小;
卷积,采用3*3卷积核,输出256个特征图,采用Relu激活函数,输出128*128大小;
上采样,采用2*2上采样,输出256*256;
卷积,采用2*2卷积核,输出128个特征图,采用Relu激活函数,输出 256*256大小,记为up10;
拼接,拼接conv2和up10,输出256个特征图,256*256大小;
卷积,采用3*3卷积核,输出128个特征图,采用Relu激活函数,输出256*256大小;
卷积,采用3*3卷积核,输出128个特征图,采用Relu激活函数,输出256*256大小;
上采样,采用2*2上采样,输出512*512;
卷积,采用2*2卷积核,输出64个特征图,采用Relu激活函数,输出512*512大小,记为up11;
拼接,拼接conv1和up11,输出128个特征图,512*512大小;
卷积,采用3*3卷积核,输出64个特征图,采用Relu激活函数,输出512*512大小;
卷积,采用3*3卷积核,输出64个特征图,采用Relu激活函数,输出512*512大小;
卷积,采用3*3卷积核,输出2个特征图,采用Relu激活函数,输出512*512大小;
卷积,采用3*3卷积核,输出1个特征图,采用sigmoid激活函数,输出512*512大小。
接收模块140,用于接收待进行肝脏癌变位置定位的CT切片图像。接收到CT切片图像后,为了增强对比,凸显肝脏组织,根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,对接收到的CT切片图像进行像素过滤,以生成过滤图像,同时保证该过滤图像的图像尺寸与接收到的CT切片图像尺寸一致。接着,对该过滤图像进行直方图均衡化处理,生成均衡化处理后的图像。最后,将均衡化处理后的图像输入到识别模型中进行定位识别。
识别模块150,用于利用预先训练的识别模型对该CT切片图像进行肝脏癌变位置的定位识别。若识别出肝脏癌变位置时,在该CT切片图像的肝脏癌变位置标注预设形式的标签。例如,在某CT切片图像上识别出某位置患有肝癌,则在识别出的肝脏癌病变区域位置生成曲线线框,并在线框内标注。
如图3所示,是本申请肝脏癌变定位方法较佳实施例的流程图。
在本实施例中,处理器12执行存储器11中存储的肝脏癌变定位程序10的计算机程序时实现肝脏癌变定位方法包括:步骤S10-步骤S50:
步骤S10,样本处理模块110获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线,每个CT切片样本图像对应标有非癌症标记或者癌症标记。例如,获取10000个CT切片样本图像,其中8000个CT切片样本图像上带有肝脏癌病变区域,2000个CT切片样本图像上未带有肝脏癌病变区域。所述病变标志点是指病变区域和非病变区域的分界点。接着,分别对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像。所述预处理具体包括:根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,如肝脏组织的灰度范围为[-100~400],分别对各个CT切片样本图像进行预设灰度范围的像素过滤,以生成对应的过滤图像,并保证各个过滤图像的图像尺寸与对应的CT切片样本图像的图像尺寸一致。接着,分别对各个过滤图像进行直方图均衡化处理,生成均衡化处理后的图像,各个均衡化处理后的图像即为预处理图像。进一步地,还可以根据直方图拉伸等方法增强对比度。
步骤S20,形变模块120分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合。其中,所述预设的形变规则为包括:对待形变处理的预处理图像,增加该预处理图像增加高斯噪声,生成对应的加噪图像。所述高斯噪声完全由其时变平均值和两瞬时平均值的协方差函数确定。例如,随机生成高斯分布的随机数,将该随机数与该预处理图像的像素值相加,将相加的值压缩到[0~225]区间内,得到对应的加噪图像。接着,在预设角度范围内,对该加噪图像进行角度随机旋转,生成对应的旋转图像。假设,预设角度范围为[-30~30],在该预设角度范围内随机选择某一角度,对该加噪图像进行该角度的旋转,生成对应的旋转图像。最后,根据预设的弹性变换规则,对该旋转图像进行弹性变换,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合。
其中,所述预设的弹性变换规则包括:针对一个旋转图像,分别对该旋转图像上的每个像素点(xi,yi)在范围为[-1~1]之间产生2个随机数Δx(xi,yi)和Δy(xi,yi),将随机数Δx(xi,yi)存放到与该旋转图像相同大小的像 素矩阵D和E的像素(xi,yi)的xi上,表示像素点(xi,yi)的x方向的移动距离,并将随机数Δy(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的yi上,表示像素点(xi,yi)的y方向的移动距离,得到2个随机数矩阵D1和E1。但应理解的是,所述范围包括但不限于[-1~1]。然后,随机生成一个以第一预设值为均值,以第二预设值为标准差的预设大小为105*105的高斯核,将该高斯核分别与随机数矩阵D1和E1分别卷积,生成2个卷积结果图像,分别为A(xi,yi)、B(xi,yi)。最后,将2个卷积结果图像作用于原图像:将该旋转图像的位置(xi,yi)的像素放到新图(xi+A(xi,yi),yi+B(xi,yi))位置上,以将所有像素移动之后得到最后的形变图像。
步骤S30,训练模块130利用图像集合中的图像对识别模型进行训练。如图4所示,是本申请识别模型训练的流程图。所述识别模型的训练步骤如下:
将所有待训练图像集合分为第一比例的训练集、第二比例的验证集。例如,将所有待训练图像集合按照7:3的比例随机分成训练集和验证集,训练集占所有待训练图像集合的70%,剩余30%的待训练图像集合作为验证集对模型进行检测。
利用训练集中的各个图像进行模型训练,以生成所述识别模型,并利用验证集中的各个图像对生成的所述识别模型进行验证。例如,将训练集中的7000个图像集合对模型进行训练,并利用验证集中的3000个图像集合进行验证,以生成最优的识别模型。
若验证通过率大于或者等于预设阈值,则训练完成,若验证通过率小于预设阈值,则增加第二预设数量的CT切片样本图像,并对增加的CT切片样本图像进行预处理及形变处理,之后流程返回将图像集合分为训练集和验证集的步骤。假设,预设阈值为98%,将验证集中的图像集合代入识别模型进行验证,若通过率大于或等于98%,则该识别模型为最优模型。若通过率小于98%,则增加2000个CT切片样本图像,并对增加的CT切片样本图像进行预处理及形变处理,流程返回到将图像集合分为训练集和验证集的步骤。
步骤S40,接收模块140接收待进行肝脏癌变位置定位的CT切片图像。接收到CT切片图像后,为了增强对比,凸显肝脏组织,根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,对接收到的CT切片图像进行像素过滤,以生成过滤图像。例如,设置肝脏的灰度值在灰度范围[-100~400]内,并对CT 切片图像进行过滤。同时保证该过滤图像的图像尺寸与接收到的CT切片图像尺寸一致。接着,对该过滤图像进行直方图均衡化处理,扩大像素个数多的灰度级,扩展图像元素取值的动态范围,生成均衡化处理后的图像,凸显肝脏与其他组织在图像中的对比度。例如,将过滤图像进行直方图均衡化处理。最后,将均衡化处理后的图像输入到识别模型中进行定位识别。
步骤S50,识别模块150利用预先训练的识别模型对该CT切片图像进行肝脏癌变位置的定位识别。当识别出肝脏癌变位置时,标注模块130在该CT切片图像的肝脏癌变位置标注预设形式的标签。例如,在某CT切片图像上识别出某位置患有肝癌,则在生成曲线线框的肝脏癌病变区域位置标注:“该CT切片图像对应的患者此处的肝脏发生癌变”。
本实施例提出的肝脏癌变定位方法,通过利用预先训练的识别模型对接收到的CT切片图像的肝脏癌变位置进行快速、精准的定位检测,提高肝脏癌变定位速度及准确率。
此外,本申请实施例还提出一种计算机可读存储介质,所述计算机可读存储介质中包括肝脏癌变定位程序10,所述肝脏癌变定位程序10被处理器执行时实现如下操作:
样本处理步骤:获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线,每个CT切片样本图像对应标有非癌症标记或者癌症标记,并对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像;
形变步骤:分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合;
训练步骤:利用图像集合中的图像对识别模型进行训练;
接收步骤:接收待进行肝脏癌变位置定位的CT切片图像;
识别步骤:将该CT切片图像输入训练好的识别模型进行肝脏癌变位置的定位识别。
优选地,所述预先训练的识别模型的训练步骤如下:
将所有待训练图像集合分为第一比例的训练集、第二比例的验证集;
利用训练集中的各个图像进行模型训练,以生成所述识别模型,并利用验证集中的各个图像对生成的所述识别模型进行验证;
若验证通过率大于或者等于预设阈值,则训练完成,若验证通过率小于预设阈值,则增加第二预设数量的CT切片样本图像,并对增加的CT切片样本图像进行预处理及形变处理,之后流程返回将图像集合分为训练集和验证集的步骤。
优选地,所述预处理包括:
根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,分别对各个CT切片样本图像进行预设灰度范围的像素过滤,以生成对应的过滤图像,并保证各个过滤图像的图像尺寸与对应的CT切片样本图像的图像尺寸一致;
分别对各个过滤图像进行直方图均衡化处理,生成均衡化处理后的图像,各个均衡化处理后的图像即为预处理图像。
优选地,所述预设的形变规则为:
增加预处理图像的高斯噪声,生成对应的加噪图像;
在预设角度范围内,对该加噪图像进行角度随机旋转,生成对应的旋转图像;
根据预设的弹性变换规则,对该旋转图像进行弹性变换,生成对应的形变图像。
优选地,所述预设的弹性变换规则为:
针对一个旋转图像,分别对该旋转图像上的每个像素点(xi,yi)在范围为[-1~1]之间产生2个随机数Δx(xi,yi)和Δy(xi,yi),将随机数Δx(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的xi上,表示像素点(xi,yi)的x方向的移动距离,并将随机数Δy(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的yi上,表示像素点(xi,yi)的y方向的移动距离,得到2个随机数矩阵D1和E1;
随机生成一个以第一预设值为均值,以第二预设值为标准差的预设大小为105*105的高斯核,将该高斯核分别与随机数矩阵D1和E1分别卷积,生成2个卷积结果图像,分别为A(xi,yi)、B(xi,yi);
将2个卷积结果图像作用于原图像:将该旋转图像的位置(xi,yi)的像素放到新图(xi+A(xi,yi),yi+B(xi,yi))位置上,以将所有像素移动之 后得到最后的形变图像。
优选地,所述接收步骤包括:
根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,利用预设灰度范围的像素对接收到的CT切片图像进行过滤,生成过滤图像,并保证该过滤图像的图像尺寸与该CT切片图像的图像尺寸一致;
对该过滤图像进行直方图均衡化处理,生成均衡化处理后的图像。
本申请之计算机可读存储介质的具体实施方式与上述肝脏癌变定位方法的具体实施方式大致相同,在此不再赘述。
上述本申请实施例序号仅仅为了描述,不代表实施例的优劣。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。基于这样的理解,本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来,该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中,包括若干指令用以使得一台终端设备(可以是手机,计算机,服务器,或者网络设备等)执行本申请各个实施例所述的方法。
以上仅为本申请的优选实施例,并非因此限制本申请的专利范围,凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换,或直接或间接运用在其他相关的技术领域,均同理包括在本申请的专利保护范围内。
Claims (20)
- 一种肝脏癌变定位方法,其特征在于,所述方法包括:样本处理步骤:获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线,每个CT切片样本图像对应标有非癌症标记或者癌症标记,并对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像;形变步骤:分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合;训练步骤:利用图像集合中的图像对识别模型进行训练;接收步骤:接收待进行肝脏癌变位置定位的CT切片图像;识别步骤:将该CT切片图像输入训练好的识别模型进行肝脏癌变位置的定位识别。
- 根据权利要求1所述的肝脏癌变定位方法,其特征在于,所述预先训练的识别模型的训练步骤如下:将所有待训练图像集合分为第一比例的训练集、第二比例的验证集;利用训练集中的各个图像进行模型训练,以生成所述识别模型,并利用验证集中的各个图像对生成的所述识别模型进行验证;若验证通过率大于或者等于预设阈值,则训练完成,若验证通过率小于预设阈值,则增加第二预设数量的CT切片样本图像,并对增加的CT切片样本图像进行预处理及形变处理,之后流程返回将图像集合分为训练集和验证集的步骤。
- 根据权利要求2所述的肝脏癌变定位方法,其特征在于,所述预先训练的识别模型为卷积神经网络模型,所述卷积神经网络模型在上采样步骤叠加卷积步骤中相同维度的特征,经过压缩空间的卷积操作对图像进行压缩,得到与叠加前相同特征空间的图像。
- 根据权利要求1所述的肝脏癌变定位方法,其特征在于,所述预处理包括:根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,分别对各个CT切片样本图像进行预设灰度范围的像素过滤,以生成对应的过滤图像,并 保证各个过滤图像的图像尺寸与对应的CT切片样本图像的图像尺寸一致;分别对各个过滤图像进行直方图均衡化处理,生成均衡化处理后的图像,各个均衡化处理后的图像即为预处理图像。
- 根据权利要求1所述的肝脏癌变定位方法,其特征在于,所述预设的形变规则为:增加预处理图像的高斯噪声,生成对应的加噪图像;在预设角度范围内,对该加噪图像进行角度随机旋转,生成对应的旋转图像;根据预设的弹性变换规则,对该旋转图像进行弹性变换,生成对应的形变图像。
- 根据权利要求5所述的肝脏癌变定位方法,其特征在于,所述预设的弹性变换规则为:针对一个旋转图像,分别对该旋转图像上的每个像素点(xi,yi)在范围为[-1~1]之间产生2个随机数Δx(xi,yi)和Δy(xi,yi),将随机数Δx(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的xi上,表示像素点(xi,yi)的x方向的移动距离,并将随机数Δy(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的yi上,表示像素点(xi,yi)的y方向的移动距离,得到2个随机数矩阵D1和E1;随机生成一个以第一预设值为均值,以第二预设值为标准差的预设大小为105*105的高斯核,将该高斯核分别与随机数矩阵D1和E1分别卷积,生成2个卷积结果图像,分别为A(xi,yi)、B(xi,yi);将2个卷积结果图像作用于原图像:将该旋转图像的位置(xi,yi)的像素放到新图(xi+A(xi,yi),yi+B(xi,yi))位置上,以将所有像素移动之后得到最后的形变图像。
- 根据权利要求1所述的肝脏癌变定位方法,其特征在于,所述接收步骤包括:根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,利用预设灰度范围的像素对接收到的CT切片图像进行过滤,生成过滤图像,并保证该过滤图像的图像尺寸与该CT切片图像的图像尺寸一致;对该过滤图像进行直方图均衡化处理,生成均衡化处理后的图像。
- 一种电子装置,其特征在于,所述装置包括:存储器、处理器,所述存储器上存储有肝脏癌变定位程序,所述肝脏癌变定位程序被所述处理器执行,可实现如下步骤:样本处理步骤:获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线,每个CT切片样本图像对应标有非癌症标记或者癌症标记,并对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像;形变步骤:分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合;训练步骤:利用图像集合中的图像对识别模型进行训练;接收步骤:接收待进行肝脏癌变位置定位的CT切片图像;识别步骤:将该CT切片图像输入训练好的识别模型进行肝脏癌变位置的定位识别。
- 根据权利要求8所述的电子装置,其特征在于,所述预先训练的识别模型的训练步骤如下:将所有待训练图像集合分为第一比例的训练集、第二比例的验证集;利用训练集中的各个图像进行模型训练,以生成所述识别模型,并利用验证集中的各个图像对生成的所述识别模型进行验证;若验证通过率大于或者等于预设阈值,则训练完成,若验证通过率小于预设阈值,则增加第二预设数量的CT切片样本图像,并对增加的CT切片样本图像进行预处理及形变处理,之后流程返回将图像集合分为训练集和验证集的步骤。
- 根据权利要求9所述的电子装置,其特征在于,所述预先训练的识别模型为卷积神经网络模型,所述卷积神经网络模型在上采样步骤叠加卷积步骤中相同维度的特征,经过压缩空间的卷积操作对图像进行压缩,得到与叠加前相同特征空间的图像。
- 根据权利要求8所述的电子装置,其特征在于,所述预处理包括:根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,分别对各个CT切片样本图像进行预设灰度范围的像素过滤,以生成对应的过滤图像,并 保证各个过滤图像的图像尺寸与对应的CT切片样本图像的图像尺寸一致;分别对各个过滤图像进行直方图均衡化处理,生成均衡化处理后的图像,各个均衡化处理后的图像即为预处理图像。
- 根据权利要求8所述的电子装置,其特征在于,所述预设的形变规则为:增加预处理图像的高斯噪声,生成对应的加噪图像;在预设角度范围内,对该加噪图像进行角度随机旋转,生成对应的旋转图像;根据预设的弹性变换规则,对该旋转图像进行弹性变换,生成对应的形变图像。
- 根据权利要求12所述的电子装置,其特征在于,所述预设的弹性变换规则为:针对一个旋转图像,分别对该旋转图像上的每个像素点(xi,yi)在范围为[-1~1]之间产生2个随机数Δx(xi,yi)和Δy(xi,yi),将随机数Δx(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的xi上,表示像素点(xi,yi)的x方向的移动距离,并将随机数Δy(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的yi上,表示像素点(xi,yi)的y方向的移动距离,得到2个随机数矩阵D1和E1;随机生成一个以第一预设值为均值,以第二预设值为标准差的预设大小为105*105的高斯核,将该高斯核分别与随机数矩阵D1和E1分别卷积,生成2个卷积结果图像,分别为A(xi,yi)、B(xi,yi);将2个卷积结果图像作用于原图像:将该旋转图像的位置(xi,yi)的像素放到新图(xi+A(xi,yi),yi+B(xi,yi))位置上,以将所有像素移动之后得到最后的形变图像。
- 根据权利要求8所述的电子装置,其特征在于,所述接收步骤包括:根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,利用预设灰度范围的像素对接收到的CT切片图像进行过滤,生成过滤图像,并保证该过滤图像的图像尺寸与该CT切片图像的图像尺寸一致;对该过滤图像进行直方图均衡化处理,生成均衡化处理后的图像。
- 一种计算机可读存储介质,其特征在于,所述计算机可读存储介质 中包括肝脏癌变定位程序,所述统肝脏癌变定位程序被处理器执行时,可实现如下步骤:样本处理步骤:获取第一预设数量的CT切片样本图像,每个CT切片样本图像上标注有病变标志点及由病变标志点限定出的病变形状曲线,每个CT切片样本图像对应标有非癌症标记或者癌症标记,并对获取的各个CT切片样本图像进行预处理,生成对应的预处理图像;形变步骤:分别对各个预处理图像按照预设的形变规则,生成对应的形变图像,将每个预处理图像及其对应的形变图像分别组成一个对应的待训练图像集合;训练步骤:利用图像集合中的图像对识别模型进行训练;接收步骤:接收待进行肝脏癌变位置定位的CT切片图像;识别步骤:将该CT切片图像输入训练好的识别模型进行肝脏癌变位置的定位识别。
- 根据权利要求15所述的计算机可读存储介质,其特征在于,所述预先训练的识别模型的训练步骤如下:将所有待训练图像集合分为第一比例的训练集、第二比例的验证集;利用训练集中的各个图像进行模型训练,以生成所述识别模型,并利用验证集中的各个图像对生成的所述识别模型进行验证;若验证通过率大于或者等于预设阈值,则训练完成,若验证通过率小于预设阈值,则增加第二预设数量的CT切片样本图像,并对增加的CT切片样本图像进行预处理及形变处理,之后流程返回将图像集合分为训练集和验证集的步骤。
- 根据权利要求15所述的计算机可读存储介质,其特征在于,所述预处理包括:根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,分别对各个CT切片样本图像进行预设灰度范围的像素过滤,以生成对应的过滤图像,并保证各个过滤图像的图像尺寸与对应的CT切片样本图像的图像尺寸一致;分别对各个过滤图像进行直方图均衡化处理,生成均衡化处理后的图像,各个均衡化处理后的图像即为预处理图像。
- 根据权利要求15所述的计算机可读存储介质,其特征在于,所述预 设的形变规则为:增加预处理图像的高斯噪声,生成对应的加噪图像;在预设角度范围内,对该加噪图像进行角度随机旋转,生成对应的旋转图像;根据预设的弹性变换规则,对该旋转图像进行弹性变换,生成对应的形变图像。
- 根据权利要求18所述的计算机可读存储介质,其特征在于,所述预设的弹性变换规则为:针对一个旋转图像,分别对该旋转图像上的每个像素点(xi,yi)在范围为[-1~1]之间产生2个随机数Δx(xi,yi)和Δy(xi,yi),将随机数Δx(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的xi上,表示像素点(xi,yi)的x方向的移动距离,并将随机数Δy(xi,yi)存放到与该旋转图像相同大小的像素矩阵D和E的像素(xi,yi)的yi上,表示像素点(xi,yi)的y方向的移动距离,得到2个随机数矩阵D1和E1;随机生成一个以第一预设值为均值,以第二预设值为标准差的预设大小为105*105的高斯核,将该高斯核分别与随机数矩阵D1和E1分别卷积,生成2个卷积结果图像,分别为A(xi,yi)、B(xi,yi);将2个卷积结果图像作用于原图像:将该旋转图像的位置(xi,yi)的像素放到新图(xi+A(xi,yi),yi+B(xi,yi))位置上,以将所有像素移动之后得到最后的形变图像。
- 根据权利要求15所述的计算机可读存储介质,其特征在于,所述接收步骤包括:根据预先确定的肝脏组织在CT切片图像上的预设灰度范围,利用预设灰度范围的像素对接收到的CT切片图像进行过滤,生成过滤图像,并保证该过滤图像的图像尺寸与该CT切片图像的图像尺寸一致;对该过滤图像进行直方图均衡化处理,生成均衡化处理后的图像。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810501877.7A CN108875734B (zh) | 2018-05-23 | 2018-05-23 | 肝脏癌变定位方法、装置及存储介质 |
CN201810501877.7 | 2018-05-23 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2019223147A1 true WO2019223147A1 (zh) | 2019-11-28 |
Family
ID=64333563
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2018/102133 WO2019223147A1 (zh) | 2018-05-23 | 2018-08-24 | 肝脏癌变定位方法、装置及存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN108875734B (zh) |
WO (1) | WO2019223147A1 (zh) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112435246A (zh) * | 2020-11-30 | 2021-03-02 | 武汉楚精灵医疗科技有限公司 | 窄带成像放大胃镜下胃癌的人工智能诊断方法 |
CN116309454A (zh) * | 2023-03-16 | 2023-06-23 | 首都师范大学 | 基于轻量化卷积核网络的病理图像智能识别方法及装置 |
Families Citing this family (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110443781A (zh) * | 2019-06-27 | 2019-11-12 | 杭州智团信息技术有限公司 | 一种肝脏数字病理的ai辅助诊断系统及方法 |
CN113496231B (zh) * | 2020-03-18 | 2024-06-18 | 北京京东乾石科技有限公司 | 分类模型训练方法、图像分类方法、装置、设备及介质 |
CN111950595A (zh) * | 2020-07-14 | 2020-11-17 | 十堰市太和医院(湖北医药学院附属医院) | 肝脏病灶图像处理方法、系统、存储介质、程序、终端 |
CN112001308B (zh) * | 2020-08-21 | 2022-03-15 | 四川大学 | 一种采用视频压缩技术和骨架特征的轻量级行为识别方法 |
CN112215217B (zh) * | 2020-12-03 | 2021-04-13 | 印迹信息科技(北京)有限公司 | 模拟医师阅片的数字图像识别方法及装置 |
CN112991214B (zh) * | 2021-03-18 | 2024-03-08 | 成都极米科技股份有限公司 | 图像处理方法、图像渲染方法、装置及影设备 |
CN113177955B (zh) * | 2021-05-10 | 2022-08-05 | 电子科技大学成都学院 | 一种基于改进图像分割算法的肺癌影像病变区域划分方法 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20110307427A1 (en) * | 2005-04-19 | 2011-12-15 | Steven Linke | Molecular markers predicting response to adjuvant therapy, or disease progression, in breast cancer |
CN107103187A (zh) * | 2017-04-10 | 2017-08-29 | 四川省肿瘤医院 | 基于深度学习的肺结节检测分级与管理的方法及系统 |
CN107730507A (zh) * | 2017-08-23 | 2018-02-23 | 成都信息工程大学 | 一种基于深度学习的病变区域自动分割方法 |
CN107767378A (zh) * | 2017-11-13 | 2018-03-06 | 浙江中医药大学 | 基于深度神经网络的gbm多模态磁共振图像分割方法 |
CN107784647A (zh) * | 2017-09-29 | 2018-03-09 | 华侨大学 | 基于多任务深度卷积网络的肝脏及其肿瘤分割方法及系统 |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101673340A (zh) * | 2009-08-13 | 2010-03-17 | 重庆大学 | 综合多方向多尺度与bp神经网络的人耳识别方法 |
US9047660B2 (en) * | 2012-03-01 | 2015-06-02 | Siemens Corporation | Network cycle features in relative neighborhood graphs |
CN103064046B (zh) * | 2012-12-25 | 2015-04-15 | 深圳先进技术研究院 | 一种基于稀疏采样的核磁共振成像的图像处理方法 |
CN106372390B (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肺癌自助健康云服务系统 |
CN106778829B (zh) * | 2016-11-28 | 2019-04-30 | 常熟理工学院 | 一种主动学习的肝脏损伤类别的图像检测方法 |
CN107153816B (zh) * | 2017-04-16 | 2021-03-23 | 五邑大学 | 一种用于鲁棒人脸识别的数据增强方法 |
-
2018
- 2018-05-23 CN CN201810501877.7A patent/CN108875734B/zh active Active
- 2018-08-24 WO PCT/CN2018/102133 patent/WO2019223147A1/zh active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20110307427A1 (en) * | 2005-04-19 | 2011-12-15 | Steven Linke | Molecular markers predicting response to adjuvant therapy, or disease progression, in breast cancer |
CN107103187A (zh) * | 2017-04-10 | 2017-08-29 | 四川省肿瘤医院 | 基于深度学习的肺结节检测分级与管理的方法及系统 |
CN107730507A (zh) * | 2017-08-23 | 2018-02-23 | 成都信息工程大学 | 一种基于深度学习的病变区域自动分割方法 |
CN107784647A (zh) * | 2017-09-29 | 2018-03-09 | 华侨大学 | 基于多任务深度卷积网络的肝脏及其肿瘤分割方法及系统 |
CN107767378A (zh) * | 2017-11-13 | 2018-03-06 | 浙江中医药大学 | 基于深度神经网络的gbm多模态磁共振图像分割方法 |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112435246A (zh) * | 2020-11-30 | 2021-03-02 | 武汉楚精灵医疗科技有限公司 | 窄带成像放大胃镜下胃癌的人工智能诊断方法 |
CN116309454A (zh) * | 2023-03-16 | 2023-06-23 | 首都师范大学 | 基于轻量化卷积核网络的病理图像智能识别方法及装置 |
CN116309454B (zh) * | 2023-03-16 | 2023-09-19 | 首都师范大学 | 基于轻量化卷积核网络的病理图像智能识别方法及装置 |
Also Published As
Publication number | Publication date |
---|---|
CN108875734B (zh) | 2021-07-23 |
CN108875734A (zh) | 2018-11-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2019223147A1 (zh) | 肝脏癌变定位方法、装置及存储介质 | |
CN110874594B (zh) | 基于语义分割网络的人体外表损伤检测方法及相关设备 | |
US10699103B2 (en) | Living body detecting method and apparatus, device and storage medium | |
CN108154509B (zh) | 癌症识别方法、装置及存储介质 | |
WO2019109526A1 (zh) | 人脸图像的年龄识别方法、装置及存储介质 | |
US11354797B2 (en) | Method, device, and system for testing an image | |
JP6819898B2 (ja) | 自動3d脳腫瘍セグメンテーション及び分類 | |
US9092697B2 (en) | Image recognition system and method for identifying similarities in different images | |
US11276490B2 (en) | Method and apparatus for classification of lesion based on learning data applying one or more augmentation methods in lesion information augmented patch of medical image | |
WO2019061658A1 (zh) | 眼镜定位方法、装置及存储介质 | |
WO2021000404A1 (zh) | 基于深度学习的目标检测方法及电子装置 | |
CN110110808B (zh) | 一种对图像进行目标标注的方法、装置及计算机记录介质 | |
US11783488B2 (en) | Method and device of extracting label in medical image | |
CN109523525B (zh) | 图像融合的恶性肺结节识别方法、装置、设备及存储介质 | |
US20190378607A1 (en) | System and method for patient privacy protection in medical images | |
CN108734708B (zh) | 胃癌识别方法、装置及存储介质 | |
CN110969046B (zh) | 人脸识别方法、设备及计算机可读存储介质 | |
CN112132812B (zh) | 证件校验方法、装置、电子设备及介质 | |
US20200013162A1 (en) | Apparatus and method for analyzing cephalometric image | |
CN116228787A (zh) | 图像勾画方法、装置、计算机设备、存储介质 | |
CN113077464A (zh) | 一种医学图像处理方法、医学图像识别方法及装置 | |
Xu et al. | Mammographic mass segmentation using multichannel and multiscale fully convolutional networks | |
CN108681731A (zh) | 一种甲状腺癌症超声图片自动标注方法及系统 | |
WO2020244076A1 (zh) | 人脸识别方法、装置、电子设备及存储介质 | |
JP7479031B2 (ja) | 顔解析のための画像正規化 |
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: 18920047 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: 18920047 Country of ref document: EP Kind code of ref document: A1 |