WO2022001237A1 - 鼻咽癌原发肿瘤图像自动识别方法及系统 - Google Patents

鼻咽癌原发肿瘤图像自动识别方法及系统 Download PDF

Info

Publication number
WO2022001237A1
WO2022001237A1 PCT/CN2021/083154 CN2021083154W WO2022001237A1 WO 2022001237 A1 WO2022001237 A1 WO 2022001237A1 CN 2021083154 W CN2021083154 W CN 2021083154W WO 2022001237 A1 WO2022001237 A1 WO 2022001237A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
sequence
dimensional
pixel
matrix
Prior art date
Application number
PCT/CN2021/083154
Other languages
English (en)
French (fr)
Inventor
魏军
朱德明
谢培梁
Original Assignee
广州柏视医疗科技有限公司
广州柏视数据科技有限公司
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 广州柏视医疗科技有限公司, 广州柏视数据科技有限公司 filed Critical 广州柏视医疗科技有限公司
Publication of WO2022001237A1 publication Critical patent/WO2022001237A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
    • G06T7/344—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods involving models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
    • 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
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10088—Magnetic resonance imaging [MRI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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

Definitions

  • the present invention relates to the technical field of medical image processing, and more particularly, to a method and system for automatically identifying images of primary tumors of nasopharyngeal carcinoma.
  • embodiments of the present invention provide a method and system for automatically identifying images of primary tumors of nasopharyngeal carcinoma.
  • an embodiment of the present invention provides a method for automatically identifying an image of a primary tumor of nasopharyngeal carcinoma, including:
  • the MR sequence three-dimensional image is registered with the CT three-dimensional image, so that the pixel positions in the MR sequence three-dimensional image are aligned with the pixel positions in the CT three-dimensional image, and the third three-dimensional CT image is registered.
  • a three-dimensional matrix is combined with a second three-dimensional matrix of the registered three-dimensional image of the MR sequence to obtain a first multi-channel matrix;
  • it also includes:
  • any channel from the second multi-channel matrix and determine a three-dimensional probability map based on the any channel.
  • the value of each pixel in the three-dimensional probability map is used to indicate that the pixel belongs to the The probability value of the region to which the primary tumor of nasopharyngeal carcinoma belongs;
  • the pixel points whose value is greater than the preset threshold in the three-dimensional probability map are marked as 1, otherwise, marked as 0, and the segmentation result of the primary tumor of nasopharyngeal carcinoma represented by any channel is obtained.
  • the registering the MR sequence three-dimensional image and the CT three-dimensional image specifically includes:
  • a transformation matrix and displacement are determined, and based on the transformation matrix and the displacement, a transformation operation and a translation operation are performed on the three-dimensional image of the MR sequence, so that the operation Pixel locations in the MR sequence 3D image are aligned with pixel locations in the CT 3D image.
  • the position information of all the pixels with a value of 1 in the first binarization map constitute a first pixel position list
  • the position information of the pixels with a value of 1 in the second binarization map constitute a the second list of pixel locations
  • the registration parameter when the registration energy function takes the minimum value is specifically determined by the following method:
  • the joint matrix is decomposed, and based on the decomposed result and the first mean value and the second mean value, the registration parameter when the registration energy function takes a minimum value is solved.
  • inputting the first multi-channel matrix into a semantic segmentation network model to obtain a second multi-channel matrix output by the semantic segmentation network model specifically includes:
  • the coding layer and the decoding layer are skip-connected.
  • the encoding layer includes a convolution layer and a downsampling layer
  • the decoding layer includes a convolution layer and an upsampling layer
  • the convolution layer in the encoding layer and the convolution layer in the decoding layer are specifically EvoNorms layers.
  • the method before the registering the MR sequence three-dimensional image and the CT three-dimensional image, the method further includes:
  • the values of the pixel points in the denoised CT 3D image and the pixel values in the denoised MR sequence 3D image are respectively updated, and the The values of all pixel points in the denoised CT 3D image and the denoised MR sequence 3D image are mapped to an interval of 0-255.
  • an embodiment of the present invention provides an automatic image identification system for nasopharyngeal carcinoma primary tumor images, including: a three-dimensional image acquisition module, a first multi-channel matrix determination module, and a second multi-channel matrix determination module.
  • the three-dimensional image acquisition module is used to acquire the CT three-dimensional image and the three-dimensional image of the magnetic resonance MR sequence of the subject;
  • the first multi-channel matrix determination module is configured to register the MR sequence 3D image with the CT 3D image, so that the pixel positions in the MR sequence 3D image are aligned with the pixel positions in the CT 3D image, combining the first three-dimensional matrix of the CT three-dimensional image with the second three-dimensional matrix of the registered three-dimensional image of the MR sequence to obtain a first multi-channel matrix;
  • the second multi-channel matrix determination module is configured to input the first multi-channel matrix into the semantic segmentation network model, and obtain a second multi-channel matrix output by the semantic segmentation network model, wherein different channels in the second multi-channel matrix are The segmentation results of nasopharyngeal carcinoma primary tumors with different risk levels, respectively.
  • an embodiment of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and running on the processor, the processor implementing the first program when executing the program
  • the processor implementing the first program when executing the program
  • an embodiment of the present invention provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, realizes the primary tumor of nasopharyngeal carcinoma as described in the first aspect The steps of the automatic image recognition method.
  • the embodiments of the present invention provide a method and system for automatically identifying primary tumor images of nasopharyngeal carcinoma, which are based on CT three-dimensional images of a subject and supplemented by MR sequence three-dimensional images to form a multi-modal semantic segmentation network model.
  • Input to achieve primary tumor identification of nasopharyngeal carcinoma on CT can effectively improve the quality of the input data, and learn the global information and detailed information of high-resolution images, which can effectively improve the prediction accuracy and generalization ability of the semantic segmentation network model.
  • the flexibility of input and output can effectively improve the work efficiency of medical workers.
  • FIG. 1 is a schematic flowchart of a method for automatically identifying a primary tumor image of nasopharyngeal carcinoma according to an embodiment of the present invention
  • FIG. 2 is a schematic structural diagram of a system for automatically identifying images of nasopharyngeal carcinoma primary tumor images provided by an embodiment of the present invention
  • FIG. 3 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
  • an embodiment of the present invention provides a method for automatically identifying a primary tumor image of nasopharyngeal carcinoma, including:
  • the MR sequence three-dimensional image is registered with the CT three-dimensional image, so that the pixel position in the MR sequence three-dimensional image is aligned with the pixel position in the CT three-dimensional image, and the CT three-dimensional image is registered.
  • the first three-dimensional matrix is combined with the second three-dimensional matrix of the registered MR sequence three-dimensional image to obtain a first multi-channel matrix;
  • the execution subject is a processor, which may be a local processor or a cloud processor, and the local processor may be a computer, A tablet, a smart phone, etc., which are not specifically limited in the embodiments of the present invention.
  • Step S1 is performed first.
  • the tested person refers to a patient who may have a primary tumor of nasopharyngeal carcinoma. Therefore, the position of the primary tumor of nasopharyngeal carcinoma needs to be determined by the automatic image recognition method of primary tumor of nasopharyngeal carcinoma provided in the embodiment of the present invention.
  • the magnetic resonance (Magnetic Resonance, MR) sequence three-dimensional images may specifically include four sequence three-dimensional images of T1, T2, T1C, and T1FSC.
  • step S2 is performed.
  • the MR sequence 3D image is registered with the CT 3D image, that is, each MR sequence 3D image is registered to the CT 3D image, so that the pixel positions in each MR sequence 3D image after registration are the same as the CT 3D image. Align pixel locations in a 3D image.
  • the first three-dimensional matrix of the CT three-dimensional image is combined with the second three-dimensional matrix of the registered three-dimensional image of the MR sequence to obtain a first multi-channel matrix.
  • the first three-dimensional matrix refers to a pixel matrix composed of pixel values of each pixel point in the CT three-dimensional image, and the elements in the first three-dimensional matrix correspond one-to-one with the pixel points in the CT three-dimensional image.
  • the second three-dimensional matrix is a pixel matrix formed by the pixel value of each pixel in the registered three-dimensional image of the MR sequence, and the elements in the second three-dimensional matrix correspond to the pixels in the three-dimensional image of the registered MR sequence one-to-one .
  • the first three-dimensional matrix of the CT three-dimensional image is combined with the second three-dimensional matrix of the registered three-dimensional image of the MR sequence. Specifically, the elements at the same position in the first three-dimensional matrix and the second three-dimensional matrix are combined, corresponding to Each element in the obtained first multi-channel matrix includes a plurality of channels.
  • the CT three-dimensional images and the three-dimensional images of each MR sequence have the same size and size, so the number of elements in the first three-dimensional matrix, the second three-dimensional matrix, and the first multi-channel matrix are all equal.
  • any number of sequence 3D images can be selected from the registered MR sequence 3D images, and the second 3D matrix of the MR sequence 3D image can be combined with the first 3D matrix of the CT 3D image to obtain the first 3D image.
  • Multichannel matrix If a sequence of 3D images is not selected, the second 3D matrix of the sequence of 3D images is assigned a zero matrix of the same dimension as the CT 3D matrix.
  • the multi-channel matrix is obtained by splicing the three-dimensional matrix of CT with the three-dimensional matrix of each MR sequence as the input of the model.
  • step S3 is performed.
  • the first multi-channel matrix is input into the semantic segmentation network model, and the second multi-channel matrix is output from the semantic segmentation network model.
  • the different channels in the second multi-channel matrix are independent of each other and represent nasopharyngeal carcinoma primary tumors with different risk levels respectively.
  • the segmentation result that is, the probability that each pixel belongs to the primary tumor of nasopharyngeal carcinoma, such as the segmentation of the primary tumor of nasopharyngeal carcinoma with a larger volume represents conservative treatment.
  • the semantic segmentation network model can be trained by a large number of CT 3D images of patients and multi-channel matrix samples corresponding to the 3D images of magnetic resonance MR sequences, and finally output the probability that each pixel belongs to the primary tumor of nasopharyngeal carcinoma.
  • the semantic segmentation network model can output the segmentation results of multiple different primary tumors of nasopharyngeal carcinoma, which represent different risk levels, and the appropriate segmentation results can be selected for subsequent analysis according to the needs.
  • the method for automatic identification of primary tumor images of nasopharyngeal carcinoma is based on the CT three-dimensional image of the tested person, supplemented by the MR sequence three-dimensional image, which constitutes a multi-modal input of the semantic segmentation network model, and realizes Nasopharyngeal carcinoma primary tumor identification on CT.
  • Combining CT 3D images and MR sequence 3D images can effectively improve the quality of the input data, and learn the global information and detailed information of high-resolution images, which can effectively improve the prediction accuracy and generalization ability of the semantic segmentation network model.
  • the flexibility of input and output can effectively improve the work efficiency of medical workers.
  • the method for automatically identifying images of nasopharyngeal carcinoma primary tumors further includes:
  • any channel from the second multi-channel matrix and determine a three-dimensional probability map based on the any channel.
  • the value of each pixel in the three-dimensional probability map is used to indicate that the pixel belongs to the The probability value of the region to which the primary tumor of nasopharyngeal carcinoma belongs;
  • the pixel points whose value is greater than the preset threshold in the three-dimensional probability map are marked as 1, otherwise, marked as 0, and the segmentation result of the primary tumor of nasopharyngeal carcinoma represented by any channel is obtained.
  • any channel is selected from the second multi-channel matrix, each channel corresponds to a matrix, and a three-dimensional probability map is determined according to the matrix corresponding to the selected channel, and each channel in the three-dimensional probability map is determined.
  • the value of each pixel point is used to represent the probability value that the pixel point belongs to the region to which the primary tumor of nasopharyngeal carcinoma belongs.
  • Pixels whose values are greater than the preset threshold in the three-dimensional probability map are marked as 1, otherwise marked as 0, and the segmentation result of the primary tumor of nasopharyngeal carcinoma represented by any channel is obtained.
  • the preset threshold may be 0.5, that is, the pixel points greater than 0.5 are marked as 1, otherwise, marked as 0, to obtain the final segmentation result of the primary tumor of nasopharyngeal carcinoma.
  • marking has the same meaning as assignment.
  • the method for automatically identifying images of nasopharyngeal carcinoma primary tumors provided in the embodiments of the present invention, wherein the three-dimensional image of the MR sequence is registered with the three-dimensional CT image, specifically includes:
  • a transformation matrix and displacement are determined, and based on the transformation matrix and the displacement, a transformation operation and a translation operation are performed on the three-dimensional image of the MR sequence, so that the operation Pixel locations in the MR sequence 3D image are aligned with pixel locations in the CT 3D image.
  • binarized images are respectively extracted from CT 3D images and MR sequence 3D images that need to be registered, that is, the value of the body part is 1, and the background value is 1. is 0, the first three-dimensional matrix and the second three-dimensional matrix are obtained respectively.
  • the first binarization map of the CT three-dimensional image and the second binarization map of the MR sequence three-dimensional image are extracted respectively, and the pixel points with a value of 1 are outputted as a list, and each element in the list is the coordinate position of the pixel point.
  • the position information of all pixels with a value of 1 in the first binarization map constitute a first pixel position list
  • the position information of pixels with a value of 1 in the second binarization map constitute a second pixel position list.
  • the registration energy function is constructed based on the position information of the pixel with a value of 1 in the first binarization map and the position information of the pixel with a value of 1 in the second binarization map, which can be specifically obtained as follows Registration energy function:
  • E is the energy function of registration
  • position information P i is the i th pixel of a second pixel position in the list
  • position information Q i is the i-th pixel of the first pixel position in the list
  • the modulus of , R and t are the registration parameters that need to be solved to make E reach the minimum value, which are the transformation matrix and displacement, respectively.
  • the transformation operation and translation operation are performed on the MR sequence 3D image, so that the pixel position in the MR sequence 3D image after the operation is aligned with the pixel position in the CT 3D image.
  • the registration parameter when the registration energy function takes the minimum value is specifically determined by the following method:
  • the joint matrix is decomposed, and based on the decomposed result and the first mean value and the second mean value, the registration parameter when the registration energy function takes a minimum value is solved.
  • the solution process of the registration energy function is as follows:
  • n is the number of pixels in the first pixel position list and the second pixel position list
  • ⁇ p is the second mean of all elements in the second pixel position list, that is, the centroid
  • ⁇ q is the first pixel position list.
  • the first mean of the elements, that is, the centroid, U, ⁇ , V are the three matrices obtained by SVD decomposition of the joint w, where ⁇ is the diagonal matrix, R * and t * are the maximum values that make E reach the minimum value, respectively. optimal solution.
  • the pixel positions in the MR sequence 3D image can be aligned with the pixel positions in the CT 3D image, thereby obtaining the registered MR sequence 3D image.
  • the first multi-channel matrix is input into the semantic segmentation network model, and the result obtained by the semantic segmentation network is obtained.
  • the second multi-channel matrix output by the model including:
  • the coding layer and the decoding layer are skip-connected.
  • the semantic segmentation network model adopted in the embodiments of the present invention includes an encoding layer and a decoding layer, that is, an encoder (encoder) and a decoder (decoder).
  • the encoding layer includes a convolution layer and a downsampling layer.
  • the encoding layer performs high-level abstract feature extraction on each channel in the input first multi-channel matrix through the convolution layer and the downsampling layer, and encodes the image corresponding to each channel. For the feature map whose size is only 1/16 of the original image, the feature map is represented in the form of a feature matrix.
  • the decoding layer includes a convolution layer and an up-sampling layer.
  • the decoding layer converts the feature map output by the encoding layer into a three-dimensional image with the same size as the original image through the convolution layer and the up-sampling layer. Probability of the region to which the primary tumor of nasopharyngeal carcinoma belongs.
  • the high-resolution feature of the shallower layer in the encoding layer is directly connected with the low-resolution feature of the higher layer in the decoding layer, which can solve the problem of loss of detailed (high-resolution) information in the high-level feature. .
  • the encoding layer includes a convolution layer and a down-sampling layer
  • the decoding layer includes a convolution layer and an up-sampling layer Floor
  • the convolution layer in the encoding layer and the convolution layer in the decoding layer are specifically EvoNorms layers.
  • the convolution layer in the encoding layer and the convolution layer in the decoding layer are both EvoNorms layers.
  • the calculation formula of the EvoNorms layer is as follows:
  • x is the input matrix
  • the dimension is d*h*w*c
  • d is the depth (z-axis)
  • h is the height (y-axis)
  • w is the width (x-axis)
  • c is the number of channels
  • v is the training The parameters that can be learned in the process
  • ⁇ is the Sigmoid function
  • s d, h, w (x) is the standard deviation of the d*h*w matrix of each channel, that is, the calculation is performed for each channel of x
  • ⁇ is equivalent to a parameter that can be learned during training in a traditional normalization layer.
  • the method for automatic identification of primary tumor images of nasopharyngeal carcinoma provided in the embodiment of the present invention, before the registration of the MR sequence three-dimensional image and the CT three-dimensional image, further includes:
  • the values of the pixel points in the denoised CT 3D image and the pixel values in the denoised MR sequence 3D image are respectively updated, and the The values of all pixel points in the denoised CT 3D image and the denoised MR sequence 3D image are mapped to an interval of 0-255.
  • the MR sequence three-dimensional image and the CT three-dimensional image may be preprocessed respectively, so that the three-dimensional images of each subject are more unified.
  • the specific process of preprocessing is as follows:
  • the automatic identification method for nasopharyngeal carcinoma primary tumor images provided in the embodiment of the present invention has the following advantages and beneficial effects:
  • the CT three-dimensional image of the subject is used, supplemented by the three-dimensional images of the four sequences of magnetic resonance MR (T2, T1, T1C, T1FSC), and the input data is effectively improved through preprocessing, registration and other steps. quality, thereby improving the prediction effect of the semantic segmentation network model.
  • the deep learning method is used, and the network structure of the encoding layer-decoding layer and the skip connection can be used to learn the global information and detailed information of the high-resolution image, and EvoNorms is used to replace the traditional normalization structure, It can effectively improve the prediction accuracy and generalization ability of the model.
  • the user can select the MR sequence that needs to be input into the model according to the requirements, and has the flexibility of the input terminal.
  • multiple GTV segmentation results can be simultaneously output, respectively representing different risk levels.
  • the AI-assisted contour processing method is implemented in the radiation therapy planning workflow, which can effectively improve the work efficiency of medical workers. Under the background of increasing demand for radiation therapy, the lack of radiation therapy resources area is particularly attractive. At the same time, the method can be extended to all other cancer types, bringing substantial advancement to the future work flow of transforming radiation therapy.
  • the embodiment of the present invention provides an image automatic identification system for nasopharyngeal carcinoma primary tumor, including: a three-dimensional image acquisition module 21 and a first multi-channel matrix determination module 22 and a second multi-channel matrix determination module 23 .
  • the three-dimensional image acquisition module 21 is used to acquire the CT three-dimensional image and the three-dimensional image of the magnetic resonance MR sequence of the subject;
  • the first multi-channel matrix determination module 22 is configured to register the MR sequence 3D image with the CT 3D image, so that the pixel positions in the MR sequence 3D image are aligned with the pixel positions in the CT 3D image , and combine the first three-dimensional matrix of the CT three-dimensional image with the second three-dimensional matrix of the registered MR sequence three-dimensional image to obtain a first multi-channel matrix;
  • the second multi-channel matrix determination module 23 is configured to input the first multi-channel matrix into the semantic segmentation network model, and obtain a second multi-channel matrix output by the semantic segmentation network model.
  • the second multi-channel matrix is different in Channels represent the segmentation results of NPC primary tumors with different risk levels, respectively.
  • each module in the automatic identification system for nasopharyngeal carcinoma primary tumor images provided in the embodiments of the present invention are in one-to-one correspondence with the operation procedures of each step in the above method embodiments, and the achieved effects are also consistent.
  • the foregoing embodiments which are not repeated in this embodiment of the present invention.
  • an embodiment of the present invention provides an electronic device, including: a processor (processor) 301, a memory (memory) 302, a communication interface (Communications Interface) 303 and a communication bus 304; of which,
  • the processor 301 , the memory 302 , and the communication interface 303 communicate with each other through the communication bus 304 .
  • the memory 302 stores program instructions that can be executed by the processor 301, and the processor 301 is used to call the program instructions in the memory 302 to execute the automatic identification of nasopharyngeal carcinoma primary tumor images provided by the above method embodiments. method.
  • the electronic device in this embodiment may be a server, a PC, or other devices in specific implementation, as long as its structure includes the processor 301 and the communication interface 303 as shown in FIG. 3 . , a memory 302 and a communication bus 304, wherein the processor 301, the communication interface 303 and the memory 302 communicate with each other through the communication bus 304, and the processor 301 can call the logic instructions in the memory 302 to execute the above method.
  • This embodiment does not limit the specific implementation form of the electronic device.
  • the logic instructions in the memory 302 may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution.
  • the computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.
  • the aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program codes .
  • an embodiment of the present invention discloses a computer program product
  • the computer program product includes a computer program stored on a non-transitory computer-readable storage medium
  • the computer program includes program instructions, and when the program instructions are executed by a computer During execution, the computer can execute the automatic identification method for the primary tumor image of nasopharyngeal carcinoma provided by the above method embodiments.
  • the embodiments of the present invention further provide a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it is implemented to execute the functions provided by the foregoing embodiments.
  • the device embodiments described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in One place, or it can be distributed over multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.
  • each embodiment can be implemented by means of software plus a necessary general hardware platform, and certainly can also be implemented by hardware.
  • the above-mentioned technical solutions can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic A disc, an optical disc, etc., includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments or some parts of the embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

一种鼻咽癌原发肿瘤图像自动识别方法及系统，以被测者的CT三维图像为基础，辅以MR序列三维图像，构成了语义分割网络模型的一个多模态输入，实现在CT上的鼻咽癌原发肿瘤识别。结合CT三维图像与MR序列三维图像，能有效提高输入数据的质量，并学习到高分辨率图像的全局信息和细节信息，能有效提高语义分割网络模型的预测准确度和泛化能力，并具备输入端与输出端的灵活性，进而有效提高医疗工作者的工作效率。

Description

鼻咽癌原发肿瘤图像自动识别方法及系统 技术领域
本发明涉及医疗图像处理技术领域，更具体地，涉及鼻咽癌原发肿瘤图像自动识别方法及系统。
背景技术
在医学领域，精准放射治疗技术已经大大提高了癌症患者的生存率。但是，这些先进的治疗方法需要对目标肿瘤的轮廓进行准确的判断，是资源密集型的。
近年来，有研究人员通过人工智能(Artificial Intelligence，AI)算法自动描绘鼻咽癌(nasopharyngeal carcinoma，NPC)原发肿瘤(gross tumor volume，GTV)。但是无法保证得到的鼻咽癌原发肿瘤的准确性。为此，现急需提供一种鼻咽癌原发肿瘤图像自动识别方法及系统。
发明内容
为克服上述问题或者至少部分地解决上述问题，本发明实施例提供了一种鼻咽癌原发肿瘤图像自动识别方法及系统。
第一方面，本发明实施例提供了一种鼻咽癌原发肿瘤图像自动识别方法，包括：
获取被测者的CT三维图像以及磁共振MR序列三维图像；
将所述MR序列三维图像与所述CT三维图像进行配准，以使所述MR序列三维图像中的像素位置与所述CT三维图像中的像素位置对齐，并将所述CT三维图像的第一三维矩阵与配准后的MR序列三维图像的第二三维矩阵进行组合，得到第一多通道矩阵；
将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，所述第二多通道矩阵中不同通道分别表示不同风险程度的鼻咽癌原发肿瘤的分割结果。
优选地，还包括：
从所述第二多通道矩阵中选取任一通道，基于所述任一通道，确定三维概率图，所述三维概率图中的每个像素点的取值用于表示所述像素点属于所述鼻咽癌原发肿瘤的所属区域的概率值；
将所述三维概率图中取值大于预设阈值的像素点标记为1，否则标记为0，得到所述任一通道表示的鼻咽癌原发肿瘤的分割结果。
优选地，所述将所述MR序列三维图像与所述CT三维图像进行配准，具体包括：
分别确定所述CT三维图像的第一二值化图以及所述MR序列三维图像的第二二值化图，并基于所述第一二值化图中取值为1的像素点的位置信息以及所述第二二值化图中取值为1的像素点的位置信息构造配准能量函数；
基于所述配准能量函数取最小值时的配准参数，确定变换矩阵和位移，基于所述变换矩阵以及所述位移，对所述MR序列三维图像进行变换操作以及平移操作，使操作后的MR序列三维图像中的像素位置与所述CT三维图像中的像素位置对齐。
优选地，所述第一二值化图中所有取值为1的像素点的位置信息构成第一像素位置列表，所述第二二值化图中取值为1的像素点的位置信息构成第二像素位置列表；
相应地，所述配准能量函数取最小值时的配准参数具体通过如下方法确定：
确定所述第一像素位置列表中所有元素的第一均值以及所述第二像素位置列表中所有元素的第二均值；
基于所述第一像素位置列表中每一元素、所述第一均值、所述第二像素列表中每一元素以及所述第二均值，构建联合矩阵；
对所述联合矩阵进行分解，并基于分解的结果以及所述第一均值和所述第二均值，求解所述配准能量函数取最小值时的配准参数。
优选地，所述将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，具体包括：
将所述第一多通道矩阵输入至所述语义分割网络模型的编码层，得到由所述编码层输出的特征矩阵；
将所述特征矩阵输入至所述语义分割网络模型的解码层，得到由所述解码层输出的第二多通道矩阵；
其中，所述编码层与所述解码层跳跃连接。
优选地，所述编码层包括卷积层和下采样层，所述解码层包括卷积层和上采样层；
所述编码层中的卷积层以及所述解码层中的卷积层具体为EvoNorms层。
优选地，所述将所述MR序列三维图像与所述CT三维图像进行配准之前，还包括：
分别对所述CT三维图像以及所述MR序列三维图像进行去噪处理，并分别计算去噪处理后的CT三维图像以及MR序列三维图像中像素点的最大取值和最小取值；
基于所述最大取值和所述最小取值，分别对去噪处理后的CT三维图像中像素点的取值以及去噪处理后的MR序列三维图像中像素点的取值进行更新，并将去噪处理后的CT三维图像中所有像素点的取值以及去噪处理后的MR序列三维图像中所有像素点的取值映射至0-255区间内。
第二方面，本发明实施例提供了一种鼻咽癌原发肿瘤图像自动识别系统，包括：三维图像获取模块、第一多通道矩阵确定模块和第二多通道矩阵确定模块。其中，
三维图像获取模块用于获取被测者的CT三维图像以及磁共振MR序列三维图像；
第一多通道矩阵确定模块用于将所述MR序列三维图像与所述CT三维图像进行配准，以使所述MR序列三维图像中的像素位置与所述CT三维图像中的像素位置对齐，并将所述CT三维图像的第一三维矩阵与配准后的MR序列三维图像的第二三维矩阵进行组合，得到第一多通道矩阵；
第二多通道矩阵确定模块用于将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，所述第二多通道矩阵中不同通道分别表示不同风险程度的鼻咽癌原发肿瘤的分割结果。
第三方面，本发明实施例提供了一种电子设备，包括：存储器、处理器以及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述程序时实现如第一方面所述的鼻咽癌原发肿瘤图像自动识别方法的步骤。
第四方面，本发明实施例提供了一种非暂态计算机可读存储介质，其上存储有计算机程序，该计算机程序被处理器执行时实现如第一方面所述的鼻咽癌原发肿瘤图像自动识别方法的步骤。
本发明实施例提供的一种鼻咽癌原发肿瘤图像自动识别方法及系统，以被测者的CT三维图像为基础，辅以MR序列三维图像，构成了语义分割网络模型的一个多模态输入，实现在CT上的鼻咽癌原发肿瘤识别。结合CT三维图像与MR序列三维图像，能有效提高输入数据的质量，并学习到高分辨率图像的全局信息和细节信息，能有效提高语义分割网络模型的预测准确度和泛化能力，并具备输入端与输出端的灵活性，进而有效提高医疗工作者的工作效率。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作一简单地介绍，显 而易见地，下面描述中的附图是本发明的一些实施例，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为本发明实施例提供的一种鼻咽癌原发肿瘤图像自动识别方法的流程示意图；
图2为本发明实施例提供的一种鼻咽癌原发肿瘤图像自动识别系统的结构示意图；
图3为本发明实施例提供的一种电子设备的结构示意图。
具体实施方式
为使本发明实施例的目的、技术方案和优点更加清楚，下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
如图1所示，本发明实施例提供了一种鼻咽癌原发肿瘤图像自动识别方法，包括：
S1，获取被测者的CT三维图像以及磁共振MR序列三维图像；
S2，将所述MR序列三维图像与所述CT三维图像进行配准，以使所述MR序列三维图像中的像素位置与所述CT三维图像中的像素位置对齐，并将所述CT三维图像的第一三维矩阵与配准后的MR序列三维图像的第二三维矩阵进行组合，得到第一多通道矩阵；
S3，将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，所述第二多通道矩阵中不同通道分别表示不同风险程度的鼻咽癌原发肿瘤的分割结果。
具体地，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方 法，其执行主体为处理器，具体可以是本地处理器，也可以是云端处理器，本地处理器具体可以是电脑、平板以及智能手机等，本发明实施例中对此不作具体限定。
首先执行步骤S1。其中，被测者是指患者，可能患有鼻咽癌原发肿瘤，因此需要借助本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法确定鼻咽癌原发肿瘤的位置。磁共振(Magnetic Resonance，MR)序列三维图像具体可以包括T1、T2、T1C、T1FSC这四个序列三维图像。
然后执行步骤S2。其中，将MR序列三维图像与CT三维图像进行配准，即是将每个MR序列三维图像配准到CT三维图像上，使配准后的每个MR序列三维图像中的像素位置均与CT三维图像中的像素位置对齐。然后将CT三维图像的第一三维矩阵与配准后的MR序列三维图像的第二三维矩阵进行组合，得到第一多通道矩阵。第一三维矩阵是指由CT三维图像中每一像素点的像素值构成的像素矩阵，第一三维矩阵中的元素与CT三维图像中的像素点一一对应。第二三维矩阵是由配准后的MR序列三维图像中每一像素点的像素值构成的像素矩阵，第二三维矩阵中的元素与配准后的MR序列三维图像中的像素点一一对应。将所述CT三维图像的第一三维矩阵与配准后的MR序列三维图像的第二三维矩阵进行组合，具体可以是第一三维矩阵和第二三维矩阵中相同位置上的元素进行组合，对应得到的第一多通道矩阵中每个元素均包含有多个通道。本发明实施例中，CT三维图像以及每个MR序列三维图像的大小尺寸均相同，因此第一三维矩阵、第二三维矩阵和第一多通道矩阵中的元素数量均相等。
需要说明的是，在配准后，可以从配准后的MR序列三维图像中选取任意若干个序列三维图像，将其第二三维矩阵与CT三维图像的第一三维矩阵进行组合，得到第一多通道矩阵。如果某个序列三维图像没有被选择，则该序列三维图像的第二三维矩阵被赋值为与CT三维矩阵相同维度大小的零矩阵。
把CT的三维矩阵与各个MR序列的三维矩阵进行拼接得到多通道的矩阵作为模型的输入。
最后执行步骤S3。其中，将第一多通道矩阵输入至语义分割网络模型，由语义分割网络模型输出第二多通道矩阵，第二多通道矩阵中不同通道相互独立，分别表示不同风险程度的鼻咽癌原发肿瘤的分割结果，即每个像素点属于鼻咽癌原发肿瘤的概率，如分割体积较大的鼻咽癌原发肿瘤表示保守治疗。
语义分割网络模型可以通过大量患者的CT三维图像以及磁共振MR序列三维图像对应得到的多通道矩阵样本训练得到，最终输出每个像素点属于鼻咽癌原发肿瘤的概率。语义分割网络模型能输出多个不同的鼻咽癌原发肿瘤的分割结果，分别代表不同风险程度，可根据需求挑选合适的分割结果进行后续分析。
本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法，以被测者的CT三维图像为基础，辅以MR序列三维图像，构成了语义分割网络模型的一个多模态输入，实现在CT上的鼻咽癌原发肿瘤识别。结合CT三维图像与MR序列三维图像，能有效提高输入数据的质量，并学习到高分辨率图像的全局信息和细节信息，能有效提高语义分割网络模型的预测准确度和泛化能力，并具备输入端与输出端的灵活性，进而有效提高医疗工作者的工作效率。
在上述实施例的基础上，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法，还包括：
从所述第二多通道矩阵中选取任一通道，基于所述任一通道，确定三维概率图，所述三维概率图中的每个像素点的取值用于表示所述像素点属于所述鼻咽癌原发肿瘤的所属区域的概率值；
将所述三维概率图中取值大于预设阈值的像素点标记为1，否则标记为0，得到所述任一通道表示的鼻咽癌原发肿瘤的分割结果。
具体地，本发明实施例中，从所述第二多通道矩阵中选取任一通道，每一通道均对应一个矩阵，根据选取的通道对应的矩阵，确定三 维概率图，三维概率图中的每个像素点的取值用于表示该像素点属于鼻咽癌原发肿瘤的所属区域的概率值。
将三维概率图中取值大于预设阈值的像素点标记为1，否则标记为0，得到所述任一通道表示的鼻咽癌原发肿瘤的分割结果。预设阈值具体可以是0.5，即将大于0.5的像素点标记为1，否则标记为0，得到最终的鼻咽癌原发肿瘤的分割结果。此处，标记与赋值的含义相同。
在上述实施例的基础上，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法，所述将所述MR序列三维图像与所述CT三维图像进行配准，具体包括：
分别确定所述CT三维图像的第一二值化图以及所述MR序列三维图像的第二二值化图，并基于所述第一二值化图中取值为1的像素点的位置信息以及所述第二二值化图中取值为1的像素点的位置信息构造配准能量函数；
基于所述配准能量函数取最小值时的配准参数，确定变换矩阵和位移，基于所述变换矩阵以及所述位移，对所述MR序列三维图像进行变换操作以及平移操作，使操作后的MR序列三维图像中的像素位置与所述CT三维图像中的像素位置对齐。
具体地，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法，对需要配准的CT三维图像以及MR序列三维图像分别提取二值化图，即身体部分的值为1，背景为0，分别得到第一三维矩阵和第二三维矩阵。分别提取CT三维图像的第一二值化图以及MR序列三维图像的第二二值化图中取值为1的像素点并输出列表，列表中的每个元素为像素点的坐标位置。第一二值化图中所有取值为1的像素点的位置信息构成第一像素位置列表，第二二值化图中取值为1的像素点的位置信息构成第二像素位置列表。
基于所述第一二值化图中取值为1的像素点的位置信息以及所述第二二值化图中取值为1的像素点的位置信息构造配准能量函数，具体可以得到如下配准能量函数：
Figure PCTCN2021083154-appb-000001
其中，E为配准能量函数，p i为第二像素位置列表中第i个像素点的位置信息，q i为第一像素位置列表中第i个像素点的位置信息，‖*‖为向量的模，R、t为需要求解的使E达到最小值的配准参数，分别为变换矩阵和位移。
基于R、t，对MR序列三维图像进行变换操作以及平移操作，使操作后的MR序列三维图像中的像素位置与CT三维图像中的像素位置对齐。
在上述实施例的基础上，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法，所述配准能量函数取最小值时的配准参数具体通过如下方法确定：
确定所述第一像素位置列表中所有元素的第一均值以及所述第二像素位置列表中所有元素的第二均值；
基于所述第一像素位置列表中每一元素、所述第一均值、所述第二像素列表中每一元素以及所述第二均值，构建联合矩阵；
对所述联合矩阵进行分解，并基于分解的结果以及所述第一均值和所述第二均值，求解所述配准能量函数取最小值时的配准参数。
具体的，本发明实施例中，配准能量函数的求解过程如下：
Figure PCTCN2021083154-appb-000002
Figure PCTCN2021083154-appb-000003
Figure PCTCN2021083154-appb-000004
W＝UΣV T
R *＝UV T
t *＝μ q-R *μ p
其中，n为第一像素位置列表以及第二像素位置列表中像素点的数量，μ p为第二像素位置列表中所有元素的第二均值，即质心，μ q为第一像素位置列表中所有元素的第一均值，也即质心，U、Σ、V为把联合w进行SVD分解后得到的三个矩阵，其中∑为对角矩阵，R *和t *分别为使E达到最小值的最优解。
通过求解得出的R *和t *，便能使MR序列三维图像中的像素位置与CT三维图像中的像素位置对齐，从而得到配准后的MR序列三维图像。
在上述实施例的基础上，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法，所述将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，具体包括：
将所述第一多通道矩阵输入至所述语义分割网络模型的编码层，得到由所述编码层输出的特征矩阵；
将所述特征矩阵输入至所述语义分割网络模型的解码层，得到由所述解码层输出的第二多通道矩阵；
其中，所述编码层与所述解码层跳跃连接。
具体地，本发明实施例中采用的语义分割网络模型包括编码层和解码层，也即编码器(encoder)和解码器(decoder)。编码层包括卷积层和下采样层，编码层通过卷积层和下采样层对输入的第一多通道矩阵中的每一通道均进行高层次抽象特征提取，将每一通道对应的图像编码为尺寸只有原图1/16的特征图，该特征图通过特征矩阵的形式表示。解码层包括卷积层和上采样层，解码层通过卷积层和上采样层将编码层输出的特征图转换为与原图尺寸相同的三维图像，其像素点的取值表示该像素点属于鼻咽癌原发肿瘤所属区域的概率。
需要说明的是，本发明实施例中编码层中较浅层的高分辨率特征与解码层中较高层的低分辨率特征直接相连，可以解决高层特征中细节(高分辨率)信息丢失的问题。
在上述实施例的基础上，本发明实施例中提供的鼻咽癌原发肿瘤 图像自动识别方法，所述编码层包括卷积层和下采样层，所述解码层包括卷积层和上采样层；
所述编码层中的卷积层以及所述解码层中的卷积层具体为EvoNorms层。
具体地，本发明实施例中，编码层中的卷积层以及解码层中的卷积层均为EvoNorms层。EvoNorms层的计算公式如下：
Figure PCTCN2021083154-appb-000005
其中，x为输入矩阵，维度为d*h*w*c，d为深度(z轴)，h为高度(y轴)，w为宽度(x轴)，c为通道数，v为在训练过程中可学习的参数，σ为Sigmoid函数，s d，h，w(x)为每个通道的d*h*w矩阵的标准差，即对x的每个通道分别进行该计算，γ、β相当于传统归一化层里的在训练过程中可学习的参数。
在上述实施例的基础上，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法，所述将所述MR序列三维图像与所述CT三维图像进行配准之前，还包括：
分别对所述CT三维图像以及所述MR序列三维图像进行去噪处理，并分别计算去噪处理后的CT三维图像以及MR序列三维图像中像素点的最大取值和最小取值；
基于所述最大取值和所述最小取值，分别对去噪处理后的CT三维图像中像素点的取值以及去噪处理后的MR序列三维图像中像素点的取值进行更新，并将去噪处理后的CT三维图像中所有像素点的取值以及去噪处理后的MR序列三维图像中所有像素点的取值映射至0-255区间内。
具体地，本发明实施例中，在进行配准之前，可以分别对MR序列三维图像与CT三维图像进行预处理，使各个被测者的各个三维图像更加统一。预处理的具体过程如下：
1)用中值滤波法分别对MR序列三维图像与CT三维图像进行去 噪处理；
2)计算去噪处理后的两个三维图像中像素点的最大取值和最小取值；
3)分别取最小取值和最大取值的一半作为新的取值范围对所有像素点的取值进行截断，即把大于最大取值的一半的像素点的取值设为最大取值一半。
4)把所有像素点的取值线性映射至[0,255]。
综上所述，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别方法与现有技术相比，具有如下优点和有益效果：
(1)本发明实施例中利用被测者的CT三维图像，辅以磁共振MR4个序列(T2、T1、T1C、T1FSC)的三维图像，通过预处理、配准等步骤有效提高输入数据的质量，从而提高语义分割网络模型的预测效果。
(2)本发明实施例中使用深度学习方法，利用编码层-解码层及跳跃连接的网络结构，能学习到高分辨率图像的全局信息和细节信息，使用EvoNorms代替传统的归一化结构，能有效提高模型的预测准确度和泛化能力。
(3)本发明实施例中使用深度学习方法，使用者可以根据需求选择需要输入模型的MR序列，具备输入端的灵活性。
(4)本发明实施例中使用深度学习方法，能同时输出多个GTV分割结果，分别表示不同的风险程度，使用者可以根据需求选择合适的输出结果，具备输出端的灵活性。
(5)本发明实施例中在放射治疗计划工作流程中实施AI辅助的轮廓加工方法，能有效提高医疗工作者的工作效率，在对放射治疗的需求不断增加的背景下，对缺乏放射治疗资源的区域尤其具有吸引力。同时该方法可以扩展适用于所有其他癌症类型，对未来改变放射治疗的工作流程带来实质性的推进。
如图2所示，在上述实施例的基础上，本发明实施例中提供了一 种鼻咽癌原发肿瘤图像自动识别系统，包括：三维图像获取模块21、第一多通道矩阵确定模块22和第二多通道矩阵确定模块23。其中，
三维图像获取模块21用于获取被测者的CT三维图像以及磁共振MR序列三维图像；
第一多通道矩阵确定模块22用于将所述MR序列三维图像与所述CT三维图像进行配准，以使所述MR序列三维图像中的像素位置与所述CT三维图像中的像素位置对齐，并将所述CT三维图像的第一三维矩阵与配准后的MR序列三维图像的第二三维矩阵进行组合，得到第一多通道矩阵；
第二多通道矩阵确定模块23用于将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，所述第二多通道矩阵中不同通道分别表示不同风险程度的鼻咽癌原发肿瘤的分割结果。
具体地，本发明实施例中提供的鼻咽癌原发肿瘤图像自动识别系统中各模块的作用与上述方法类实施例中各步骤的操作流程是一一对应的，实现的效果也是一致的，具体参见上述实施例，本发明实施例中对此不再赘述。
图3所示，在上述实施例的基础上，本发明实施例中提供了一种电子设备，包括：处理器(processor)301、存储器(memory)302、通信接口(Communications Interface)303和通信总线304；其中，
所述处理器301、存储器302、通信接口303通过通信总线304完成相互间的通信。所述存储器302存储有可被所述处理器301执行的程序指令，处理器301用于调用存储器302中的程序指令，以执行上述各方法实施例所提供的鼻咽癌原发肿瘤图像自动识别方法。
需要说明的是，本实施例中的电子设备在具体实现时可以为服务器，也可以为PC机，还可以为其他设备，只要其结构中包括如图3所示的处理器301、通信接口303、存储器302和通信总线304，其中处理器301、通信接口303和存储器302通过通信总线304完成相互间的 通信，且处理器301可以调用存储器302中的逻辑指令以执行上述方法即可。本实施例不对电子设备的具体实现形式进行限定。
存储器302中的逻辑指令可以通过软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案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一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本发明各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
进一步地，本发明实施例公开一种计算机程序产品，所述计算机程序产品包括存储在非暂态计算机可读存储介质上的计算机程序，所述计算机程序包括程序指令，当所述程序指令被计算机执行时，计算机能够执行上述各方法实施例所提供的鼻咽癌原发肿瘤图像自动识别方法。
在上述实施例的基础上，本发明实施例还提供一种非暂态计算机可读存储介质，其上存储有计算机程序，该计算机程序被处理器执行时实现以执行上述各实施例提供的鼻咽癌原发肿瘤图像自动识别方法。
以上所描述的装置实施例仅仅是示意性的，其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。本领域普通技术人员在不付出创造性的劳动的情况下，即可以理解并实施。
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 到各实施方式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。基于这样的理解，上述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，该计算机软件产品可以存储在计算机可读存储介质中，如ROM/RAM、磁碟、光盘等，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行各个实施例或者实施例的某些部分所述的方法。
最后应说明的是：以上实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的精神和范围。

Claims (10)

  1. 一种鼻咽癌原发肿瘤图像自动识别方法，其特征在于，包括：
    获取被测者的CT三维图像以及磁共振MR序列三维图像；
    将所述MR序列三维图像与所述CT三维图像进行配准，以使所述MR序列三维图像中的像素位置与所述CT三维图像中的像素位置对齐，并将所述CT三维图像的第一三维矩阵与配准后的MR序列三维图像的第二三维矩阵进行组合，得到第一多通道矩阵；
    将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，所述第二多通道矩阵中不同通道分别表示不同风险程度的鼻咽癌原发肿瘤的分割结果。
  2. 根据权利要求1所述的鼻咽癌原发肿瘤图像自动识别方法，其特征在于，还包括：
    从所述第二多通道矩阵中选取任一通道，基于所述任一通道，确定三维概率图，所述三维概率图中的每个像素点的取值用于表示所述像素点属于所述鼻咽癌原发肿瘤的所属区域的概率值；
    将所述三维概率图中取值大于预设阈值的像素点标记为1，否则标记为0，得到所述任一通道表示的鼻咽癌原发肿瘤的分割结果。
  3. 根据权利要求1所述的鼻咽癌原发肿瘤图像自动识别方法，其特征在于，所述将所述MR序列三维图像与所述CT三维图像进行配准，具体包括：
    分别确定所述CT三维图像的第一二值化图以及所述MR序列三维图像的第二二值化图，并基于所述第一二值化图中取值为1的像素点的位置信息以及所述第二二值化图中取值为1的像素点的位置信息构造配准能量函数；
    基于所述配准能量函数取最小值时的配准参数，确定变换矩阵和位移，基于所述变换矩阵以及所述位移，对所述MR序列三维图像进行变换操作以及平移操作，使操作后的MR序列三维图像中的像素位 置与所述CT三维图像中的像素位置对齐。
  4. 根据权利要求3所述的鼻咽癌原发肿瘤图像自动识别方法，其特征在于，所述第一二值化图中所有取值为1的像素点的位置信息构成第一像素位置列表，所述第二二值化图中取值为1的像素点的位置信息构成第二像素位置列表；
    相应地，所述配准能量函数取最小值时的配准参数具体通过如下方法确定：
    确定所述第一像素位置列表中所有元素的第一均值以及所述第二像素位置列表中所有元素的第二均值；
    基于所述第一像素位置列表中每一元素、所述第一均值、所述第二像素列表中每一元素以及所述第二均值，构建联合矩阵；
    对所述联合矩阵进行分解，并基于分解的结果以及所述第一均值和所述第二均值，求解所述配准能量函数取最小值时的配准参数。
  5. 根据权利要求1所述的鼻咽癌原发肿瘤图像自动识别方法，其特征在于，所述将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，具体包括：
    将所述第一多通道矩阵输入至所述语义分割网络模型的编码层，得到由所述编码层输出的特征矩阵；
    将所述特征矩阵输入至所述语义分割网络模型的解码层，得到由所述解码层输出的第二多通道矩阵；
    其中，所述编码层与所述解码层跳跃连接。
  6. 根据权利要求5所述的鼻咽癌原发肿瘤图像自动识别方法，其特征在于，所述编码层包括卷积层和下采样层，所述解码层包括卷积层和上采样层；
    所述编码层中的卷积层以及所述解码层中的卷积层具体为EvoNorms层。
  7. 根据权利要求1-6任一项所述的鼻咽癌原发肿瘤图像自动识别方法，其特征在于，所述将所述MR序列三维图像与所述CT三维图像 进行配准之前，还包括：
    分别对所述CT三维图像以及所述MR序列三维图像进行去噪处理，并分别计算去噪处理后的CT三维图像以及MR序列三维图像中像素点的最大取值和最小取值；
    基于所述最大取值和所述最小取值，分别对去噪处理后的CT三维图像中像素点的取值以及去噪处理后的MR序列三维图像中像素点的取值进行更新，并将去噪处理后的CT三维图像中所有像素点的取值以及去噪处理后的MR序列三维图像中所有像素点的取值映射至0-255区间内。
  8. 一种鼻咽癌原发肿瘤图像自动识别系统，其特征在于，包括：
    三维图像获取模块，用于获取被测者的CT三维图像以及磁共振MR序列三维图像；
    第一多通道矩阵确定模块，用于将所述MR序列三维图像与所述CT三维图像进行配准，以使所述MR序列三维图像中的像素位置与所述CT三维图像中的像素位置对齐，并将所述CT三维图像的第一三维矩阵与配准后的MR序列三维图像的第二三维矩阵进行组合，得到第一多通道矩阵；
    第二多通道矩阵确定模块，用于将所述第一多通道矩阵输入至语义分割网络模型，得到由所述语义分割网络模型输出的第二多通道矩阵，所述第二多通道矩阵中不同通道分别表示不同风险程度的鼻咽癌原发肿瘤的分割结果。
  9. 一种电子设备，包括：存储器、处理器以及存储在存储器上并可在处理器上运行的计算机程序，其特征在于，所述处理器执行所述程序时实现如权利要求1-7中任一项所述的鼻咽癌原发肿瘤图像自动识别方法的步骤。
  10. 一种非暂态计算机可读存储介质，其上存储有计算机程序，其特征在于，该计算机程序被处理器执行时实现如权利要求1-7中任一项所述的鼻咽癌原发肿瘤图像自动识别方法的步骤。
PCT/CN2021/083154 2020-06-28 2021-03-26 鼻咽癌原发肿瘤图像自动识别方法及系统 WO2022001237A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010595992.2 2020-06-28
CN202010595992.2A CN111784706B (zh) 2020-06-28 2020-06-28 鼻咽癌原发肿瘤图像自动识别方法及系统

Publications (1)

Publication Number Publication Date
WO2022001237A1 true WO2022001237A1 (zh) 2022-01-06

Family

ID=72760136

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/083154 WO2022001237A1 (zh) 2020-06-28 2021-03-26 鼻咽癌原发肿瘤图像自动识别方法及系统

Country Status (2)

Country Link
CN (1) CN111784706B (zh)
WO (1) WO2022001237A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115170510A (zh) * 2022-07-04 2022-10-11 北京医准智能科技有限公司 一种病灶检测方法、装置、电子设备及可读存储介质
CN117173092A (zh) * 2023-06-28 2023-12-05 中山大学肿瘤防治中心（中山大学附属肿瘤医院、中山大学肿瘤研究所） 一种基于图像处理的鼻咽癌放疗方法及系统
CN117197472A (zh) * 2023-11-07 2023-12-08 四川农业大学 基于鼻出血内窥镜影像的高效师生半监督分割方法及装置
CN117288659A (zh) * 2023-11-24 2023-12-26 深圳市前海高新国际医疗管理有限公司 基于ai的肿瘤干细胞与正常细胞的形态学分析系统

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111784706B (zh) * 2020-06-28 2021-06-04 广州柏视医疗科技有限公司 鼻咽癌原发肿瘤图像自动识别方法及系统
CN112508910A (zh) * 2020-12-02 2021-03-16 创新奇智(深圳)技术有限公司 用于多分类缺陷检测的缺陷提取方法及装置
CN113298855B (zh) * 2021-05-27 2021-12-28 广州柏视医疗科技有限公司 基于自动勾画的图像配准方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1952980A (zh) * 2006-11-17 2007-04-25 南方医科大学 三维图像的快速配准方法
US8588498B2 (en) * 2011-09-30 2013-11-19 General Electric Company System and method for segmenting bones on MR images
CN108257134A (zh) * 2017-12-21 2018-07-06 深圳大学 基于深度学习的鼻咽癌病灶自动分割方法和系统
CN110211079A (zh) * 2019-05-22 2019-09-06 首都医科大学 医学图像的融合方法及装置
CN111091560A (zh) * 2019-12-19 2020-05-01 广州柏视医疗科技有限公司 鼻咽癌原发肿瘤图像识别方法及系统
CN111784706A (zh) * 2020-06-28 2020-10-16 广州柏视医疗科技有限公司 鼻咽癌原发肿瘤图像自动识别方法及系统

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2732423A4 (en) * 2011-07-13 2014-11-26 Multiple Myeloma Res Foundation Inc METHOD FOR DETECTING AND DISTRIBUTING DATA
CN109410188B (zh) * 2017-10-13 2021-06-04 科亚医疗科技股份有限公司 用于对医学图像进行分割的系统和方法
CN109272510B (zh) * 2018-07-24 2021-06-04 清华大学 一种三维医学图像中管状结构的分割方法
CN110232691A (zh) * 2019-04-18 2019-09-13 浙江大学山东工业技术研究院 一种多模态的ct影像的分割方法
CN110464353A (zh) * 2019-08-21 2019-11-19 南方医科大学 一种基于深度卷积神经网络的伪ct合成方法及应用
CN111091589B (zh) * 2019-11-25 2023-11-17 北京理工大学 基于多尺度监督学习的超声和核磁图像配准方法及装置
CN112508965B (zh) * 2019-12-10 2023-08-22 广州柏视医疗科技有限公司 医学影像中正常器官的轮廓线自动勾画系统
CN110969619B (zh) * 2019-12-19 2021-06-29 广州柏视医疗科技有限公司 一种自动识别鼻咽癌原发肿瘤的方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1952980A (zh) * 2006-11-17 2007-04-25 南方医科大学 三维图像的快速配准方法
US8588498B2 (en) * 2011-09-30 2013-11-19 General Electric Company System and method for segmenting bones on MR images
CN108257134A (zh) * 2017-12-21 2018-07-06 深圳大学 基于深度学习的鼻咽癌病灶自动分割方法和系统
CN110211079A (zh) * 2019-05-22 2019-09-06 首都医科大学 医学图像的融合方法及装置
CN111091560A (zh) * 2019-12-19 2020-05-01 广州柏视医疗科技有限公司 鼻咽癌原发肿瘤图像识别方法及系统
CN111784706A (zh) * 2020-06-28 2020-10-16 广州柏视医疗科技有限公司 鼻咽癌原发肿瘤图像自动识别方法及系统

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115170510A (zh) * 2022-07-04 2022-10-11 北京医准智能科技有限公司 一种病灶检测方法、装置、电子设备及可读存储介质
CN117173092A (zh) * 2023-06-28 2023-12-05 中山大学肿瘤防治中心（中山大学附属肿瘤医院、中山大学肿瘤研究所） 一种基于图像处理的鼻咽癌放疗方法及系统
CN117173092B (zh) * 2023-06-28 2024-04-09 中山大学肿瘤防治中心（中山大学附属肿瘤医院、中山大学肿瘤研究所） 一种基于图像处理的鼻咽癌放疗方法及系统
CN117197472A (zh) * 2023-11-07 2023-12-08 四川农业大学 基于鼻出血内窥镜影像的高效师生半监督分割方法及装置
CN117197472B (zh) * 2023-11-07 2024-03-08 四川农业大学 基于鼻出血内窥镜影像的高效师生半监督分割方法及装置
CN117288659A (zh) * 2023-11-24 2023-12-26 深圳市前海高新国际医疗管理有限公司 基于ai的肿瘤干细胞与正常细胞的形态学分析系统
CN117288659B (zh) * 2023-11-24 2024-02-13 深圳市前海高新国际医疗管理有限公司 基于ai的肿瘤干细胞与正常细胞的形态学分析系统

Also Published As

Publication number Publication date
CN111784706B (zh) 2021-06-04
CN111784706A (zh) 2020-10-16

Similar Documents

Publication Publication Date Title
WO2022001237A1 (zh) 鼻咽癌原发肿瘤图像自动识别方法及系统
CN110298383B (zh) 基于多模态深度学习的病理分类方法及系统
CN111091589A (zh) 基于多尺度监督学习的超声和核磁图像配准方法及装置
CN113487657B (zh) 一种基于深度学习的模态转换方法
CN110909618B (zh) 一种宠物身份的识别方法及装置
CN110555835B (zh) 一种脑片图像区域划分方法及装置
CN110969619B (zh) 一种自动识别鼻咽癌原发肿瘤的方法及装置
WO2022127500A1 (zh) 基于多个神经网络的mri图像分割方法、装置及设备
CN114494296A (zh) 一种基于Unet和Transformer相融合的脑部胶质瘤分割方法与系统
WO2024011835A1 (zh) 一种图像处理方法、装置、设备及可读存储介质
CN116596846A (zh) 图像分割方法、图像分割模型构建方法、装置及介质
CN112836625A (zh) 人脸活体检测方法、装置、电子设备
CN115018809A (zh) Ct图像的靶区分割识别方法及系统
CN114972362A (zh) 一种基于RMAU-Net网络的医学图像自动分割方法与系统
CN112927246A (zh) 肺部轮廓分割、肿瘤免疫浸润分类系统及方法
CN112560710A (zh) 一种用于构建指静脉识别系统的方法及指静脉识别系统
CN111091560A (zh) 鼻咽癌原发肿瘤图像识别方法及系统
CN109829885B (zh) 一种基于深度语义分割网络的自动识别鼻咽癌原发肿瘤方法
CN108597589B (zh) 模型生成方法、目标检测方法及医学成像系统
CN110826565A (zh) 基于跨连接的卷积神经网络齿痕舌象分类方法及系统
CN111582449B (zh) 一种目标域检测网络的训练方法、装置、设备及存储介质
EP4167184A1 (en) Systems and methods for plaque identification, plaque composition analysis, and plaque stability detection
WO2022227193A1 (zh) 肝脏区域分割方法、装置、电子设备及存储介质
CN113487579B (zh) 自动勾画模型的多模态迁移方法
CN115239740A (zh) 基于GT-UNet的全心分割算法

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: 21831715

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: 21831715

Country of ref document: EP

Kind code of ref document: A1