A nasopharyngeal carcinoma lesion segmentation method inspired by clinical decision process

Abstract

This invention proposes a nasopharyngeal carcinoma (NPC) lesion segmentation method inspired by clinical decision-making processes. It addresses the problems of low segmentation efficiency and uninterpretable segmentation results in current NPC tumor region segmentation methods. First, a diagnostic model is used to diagnose preprocessed 3D MRI images, identifying cancerous 3D MRI images. Then, a lesion segmentation model is used to perform multi-scale segmentation of the cancerous 3D MRI images, progressively refining the segmentation results. This effectively improves the segmentation efficiency of NPC lesions and ensures the interpretability of the results.

Description

Technical Field

[0001] This invention relates to the technical field of nasopharyngeal carcinoma lesion segmentation, and in particular to a nasopharyngeal carcinoma lesion segmentation method inspired by clinical decision-making processes. Background Technology

[0002] Nasopharyngeal carcinoma is an epithelial malignant tumor that commonly occurs in the roof and lateral walls of the nasopharynx and is one of the most common malignant tumors in my country. During the diagnosis and treatment of nasopharyngeal carcinoma, patients often need to undergo magnetic resonance imaging (MRI) scans from the nasopharynx to the neck to accurately determine the extent of the lesion and delineate the target area, i.e., lesion segmentation. Lesion segmentation is often an intermediate part of the entire nasopharyngeal carcinoma analysis process; inaccurate segmentation results can lead to deviations in subsequent analysis, thus affecting the final diagnosis. Nasopharyngeal carcinoma staging depends on the location, number, and size of the primary lesion and metastatic lymph nodes; accurate lesion segmentation is crucial for the clinical staging and treatment of nasopharyngeal carcinoma.

[0003] Existing literature discloses a method for nasopharyngeal carcinoma (NPC) tumor region segmentation. First, an improved recall loss function is used to pre-train the NPC tumor region segmentation model to obtain the approximate location of the tumor region. Then, the Dice loss function is used to further train the residual U-Net and predict multiple regions on the image. Finally, the prediction results of all regions are combined to obtain a complete delineation of the tumor target area. However, this method has two shortcomings. First, it only focuses on delineating NPC lesions from cancerous sections. Clinicians need to classify each section in each 3D MRI image beforehand, selecting sections containing cancer cells, and then inputting them into the corresponding algorithm for lesion segmentation. This approach has low segmentation efficiency and requires a high level of clinical experience from the physician. Second, the U-Net framework used in this NPC tumor region segmentation model only splices cross-scale features through simple skip connections. The entire semantic segmentation process only outputs a two-dimensional segmentation result. The uninterpretability of this "black box" model may cause clinicians to doubt its decision-making process, thus hindering the clinical application of the model. Summary of the Invention

[0004] To address the issues of low segmentation efficiency and uninterpretable results in current nasopharyngeal carcinoma (NPC) tumor region segmentation methods, this invention proposes an NPC segmentation method inspired by clinical decision-making processes. This method effectively improves NPC segmentation efficiency and ensures the interpretability of NPC results.

[0005] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows:

[0006] A method for segmenting nasopharyngeal carcinoma lesions inspired by clinical decision-making processes, comprising the following steps:

[0007] S1. Obtain a certain number of three-dimensional MRI images of nasopharyngeal carcinoma patients and form a three-dimensional MRI image dataset;

[0008] S2. Preprocess the three-dimensional MRI images in the three-dimensional MRI image dataset;

[0009] S3. Construct a cancer diagnosis model based on three-dimensional MRI images for diagnosing cancer;

[0010] S4. Divide the 3D MRI image dataset into a training set and a test set. Use the training set to train the cancer diagnosis model and use the test set to test the effectiveness of the cancer diagnosis model to obtain a well-trained cancer diagnosis model.

[0011] S5. Input the preprocessed 3D MRI images from S2 into the trained cancer diagnosis model and output the 3D MRI images of the cancer.

[0012] S6. Construct a lesion segmentation model for multi-scale segmentation of three-dimensional MRI images of cancerous lesions;

[0013] S7. Use three-dimensional MRI images of cancerous lesions to train a lesion segmentation model to obtain a trained lesion segmentation model;

[0014] S8. Acquire the MRI image to be predicted, input the MRI image to be predicted into the trained cancer diagnosis model, output the MRI image to be predicted as cancer, input the MRI image to be predicted as cancer into the trained lesion segmentation model, and output the lesion segmentation result of the MRI image to be predicted as cancer.

[0015] Preferably, the three-dimensional MRI images of the nasopharyngeal carcinoma patient in step S1 include T1-weighted or T1-weighted contrast-enhanced images of the head and T2-weighted images of the neck.

[0016] Preferably, step S2, which preprocesses the three-dimensional MRI images in the three-dimensional MRI image dataset, involves scaling the pixel values ​​of the three-dimensional MRI images to the range [0,1] using an extremum normalization operation.

[0017] Preferably, the cancer diagnosis model in S3 consists of a feature extraction submodule, a convolutional layer containing a 1x1 convolutional kernel, a sigmoid function activation layer, and a cancer probability diagnosis submodule connected in sequence. The feature extraction submodule is composed of a residual network without fully connected layers, and the cancer probability diagnosis submodule is composed of a parallel distribution alignment layer and an adaptive pooling layer.

[0018] Preferably, the processing procedure of the cancer diagnosis model for three-dimensional MRI images is as follows:

[0019] S31. Let the height of the three-dimensional MRI image be H and the width be W. Input the three-dimensional MRI image into the feature extraction submodule of the cancer diagnosis model. Use the feature extraction submodule to extract the depth features of the three-dimensional MRI image and output the first depth feature map of the three-dimensional MRI image. Use a convolutional layer to perform a 1x1 convolution operation on the first depth feature map and output the second depth feature map. Use a sigmoid function activation layer to perform a sigmoid transformation on the second depth feature map and convert the second depth feature map into a saliency map.

[0020] S32. An expert prior distribution alignment layer is used to align the saliency map, and an adaptive pooling layer is used to adaptively select k important pixels in the saliency map. The mean of these k important pixels is calculated, and the mean value is recorded as the cancer probability corresponding to the 3D MRI image, i.e., the image-level label prediction result of the 3D MRI image.

[0021] S33. Determine the image-level label prediction results Does it exceed the diagnostic threshold? If so, then... The value is denoted as 1, indicating the image-level label prediction result. The corresponding 3D MRI image is a 3D MRI image of the cancerous lesion; otherwise, the 3D MRI image of the cancerous lesion will be output. Recorded as 0, this indicates the image-level label prediction result. The corresponding 3D MRI image is a normal 3D MRI image.

[0022] Preferably, the specific process of expert prior distribution alignment of the saliency map in the distribution alignment layer of S32 is as follows:

[0023] S321. Identify the saliency map corresponding to the three-dimensional MRI images of batch n;

[0024] S322. Calculate the probability mass distribution p of the saliency map described in S321. M p M The calculation expression is:

[0025]

[0026] Among them, M (i) Let represent the saliency map corresponding to the nth batch of 3D MRI images, and let ψ(.) regularization operation represent scaling the L1 norm of the entire matrix to 1;

[0027] S322. Determine the expert prior distribution of nasopharyngeal carcinoma in anatomical cross-sections. E The expert prior distribution p of nasopharyngeal carcinoma on anatomical cross sections. E Quantized to a Gaussian mixture distribution, p E The quantization expression is:

[0028]

[0029] in, Let π represent the i-th Gaussian distribution. i Represents the component weights and ∑π i =1, c represents the number of components, and the mean μ i The covariance Σ is determined by the anatomical location of the high-incidence sites of nasopharyngeal carcinoma. i This indicates positional perturbations caused by individual differences;

[0030] S323. Make the probability mass distribution p of the saliency graph... M Compared with expert prior distribution p E The distribution distance between them is the smallest.

[0031] Preferably, any three-dimensional MRI image in the S4 training set is matched with an image-level label y and a pixel-level label G. When y = 0, it indicates that the three-dimensional MRI image is a normal three-dimensional MRI image; when y = 1, it indicates that the three-dimensional MRI image is a cancerous three-dimensional MRI image; when G = 0, it indicates that the pixels of the three-dimensional MRI image are normal pixels; when G = 1, it indicates that the pixels of the three-dimensional MRI image are cancerous pixels.

[0032] Preferably, the lesion segmentation model described in S6 consists of a connected fully convolutional encoder and decoder. The specific processing procedure of the lesion segmentation model for the input three-dimensional MRI image of cancerous lesions is as follows:

[0033] S61. Use a fully convolutional encoder to extract multi-scale depth features from the three-dimensional MRI images of cancerous lesions, and output the multi-scale depth feature information of the three-dimensional MRI images of cancerous lesions to the decoder.

[0034] S62. The decoder decodes the multi-scale depth feature information in a progressively refined manner, obtaining the pixel-level label prediction result P of the multi-scale depth feature information. k ;

[0035] S63. Determine the pixel-level label prediction result P k If the value is greater than the segmentation threshold, the pixel-level label prediction result is recorded as 1, indicating that the pixel in the 3D MRI image corresponding to the pixel-level label prediction result is a cancerous pixel; otherwise, the pixel-level label prediction result is recorded as 0, indicating that the pixel in the 3D MRI image corresponding to the pixel-level label prediction result is a normal pixel.

[0036] S64. Based on the pixel-level label prediction result P from step S63 k The judgment result outputs the lesion segmentation result of the three-dimensional MRI image of cancer.

[0037] Preferably, when training the cancer diagnosis model, the training process uses a first loss function. As the total loss function, The expression is:

[0038]

[0039] in, Let λ represent the cross-entropy and λ represent the non-negative hyperparameter. Indicates alignment loss;

[0040] When training the lesion segmentation model, the training process uses the second loss function. and the third loss function As the total loss function, the second loss function The expression is:

[0041]

[0042] in, This represents the loss due to the mean square error.

[0043] Third loss function The expression is:

[0044]

[0045] in, G represents the cross-entropy. k Sampling from G 2 k The result is doubled.

[0046] This invention also proposes a nasopharyngeal carcinoma lesion segmentation system inspired by clinical decision-making processes, the system comprising:

[0047] The acquisition module is used to acquire a certain number of three-dimensional MRI images of nasopharyngeal carcinoma patients to form a three-dimensional MRI image dataset.

[0048] The preprocessing module is used to preprocess the three-dimensional MRI images in the three-dimensional MRI image dataset;

[0049] The cancer diagnosis model construction module is used to construct a cancer diagnosis model based on three-dimensional MRI images for diagnosing cancer.

[0050] The cancer diagnosis model training module is used to divide the 3D MRI image dataset into a training set and a test set. The training set is used to train the cancer diagnosis model, and the test set is used to test the effectiveness of the cancer diagnosis model, so as to obtain a trained cancer diagnosis model.

[0051] The output module is used to input the three-dimensional MRI images from the acquisition module into the trained cancer diagnosis model and output the three-dimensional MRI images of the cancer.

[0052] The lesion segmentation model construction module is used to construct a lesion segmentation model for multi-scale segmentation of three-dimensional MRI images of cancerous lesions.

[0053] The lesion segmentation model training module is used to train a lesion segmentation model using three-dimensional MRI images of cancerous lesions, and obtain a trained lesion segmentation model.

[0054] The prediction module is used to acquire MRI images to be predicted, input the MRI images to be predicted into a trained cancer diagnosis model, and output the MRI images of cancer to be predicted. It also inputs the MRI images of cancer to be predicted into a trained lesion segmentation model and outputs the lesion segmentation results of the MRI images of cancer to be predicted.

[0055] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0056] This invention proposes a nasopharyngeal carcinoma lesion segmentation method inspired by clinical decision-making processes. It establishes a deep learning framework that integrates cancer diagnosis using a diagnostic model with lesion segmentation using a segmentation model. First, the diagnostic model is used to diagnose pre-processed 3D MRI images, identifying cancerous 3D MRI images. The diagnostic process simulates a clinician's search for suspicious lesions starting from common nasopharyngeal carcinoma sites and drawing a final conclusion from locally suspicious areas. Then, the lesion segmentation model is used to perform multi-scale segmentation of the cancerous 3D MRI images, progressively refining the segmentation results. The segmentation process simulates a clinician's multi-scale segmentation process after selecting different fields of view on the slice. The entire process, from diagnosis to segmentation, is fully automated and does not rely on prior diagnosis by radiation oncologists, effectively improving the efficiency of nasopharyngeal carcinoma lesion segmentation and ensuring the interpretability of the results. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating a nasopharyngeal carcinoma lesion segmentation method inspired by clinical decision-making processes proposed in this embodiment of the invention.

[0058] Figure 2 A saliency map visualization diagram illustrating the diagnostic stage proposed in this embodiment of the invention;

[0059] Figure 3 A schematic diagram illustrating the multi-scale segmentation process of the lesion segmentation stage proposed in this embodiment of the invention;

[0060] Figure 4 This is a schematic diagram of the structure of the cancer diagnosis model proposed in the embodiments of the present invention;

[0061] Figure 5 This is a schematic diagram of the lesion segmentation model proposed in the embodiments of the present invention;

[0062] Figure 6 This is a schematic diagram illustrating the structure of a nasopharyngeal carcinoma lesion system inspired by clinical decision-making processes, as presented in an embodiment of the present invention. Detailed Implementation

[0063] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0064] To better illustrate this embodiment, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent actual dimensions. The descriptions of directions such as "up" and "down" are not intended to limit this patent.

[0065] It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings;

[0066] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0067] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0068] Example 1

[0069] like Figure 1 As shown, this embodiment proposes a nasopharyngeal carcinoma lesion segmentation method inspired by clinical decision-making processes, including the following steps:

[0070] S1. Obtain a certain number of three-dimensional MRI images of nasopharyngeal carcinoma patients and form a three-dimensional MRI image dataset;

[0071] In S1, the three-dimensional MRI images of the nasopharyngeal carcinoma patient include T1-weighted or T1-weighted contrast-enhanced images of the head and T2-weighted images of the neck.

[0072] S2. Preprocess the three-dimensional MRI images in the three-dimensional MRI image dataset;

[0073] In S2, the specific steps for preprocessing the three-dimensional MRI images in the three-dimensional MRI image dataset are as follows: using the maximum and minimum value normalization operation, the pixel values ​​of the three-dimensional MRI images are scaled to the range [0,1].

[0074] S3. Construct a cancer diagnosis model based on three-dimensional MRI images for diagnosing cancer;

[0075] S4. Divide the 3D MRI image dataset into a training set and a test set. Use the training set to train the cancer diagnosis model and use the test set to test the effectiveness of the cancer diagnosis model to obtain a well-trained cancer diagnosis model.

[0076] In S4, any 3D MRI image in the training set is matched with an image-level label y and a pixel-level label G. When y = 0, it indicates that the 3D MRI image is a normal 3D MRI image; when y = 1, it indicates that the 3D MRI image is a cancerous 3D MRI image; when G = 0, it indicates that the pixels of the 3D MRI image are normal pixels; when G = 1, it indicates that the pixels of the 3D MRI image are cancerous pixels.

[0077] S5. Input the preprocessed 3D MRI images from S2 into the trained cancer diagnosis model and output the 3D MRI images of the cancer.

[0078] S6. Construct a lesion segmentation model for multi-scale segmentation of three-dimensional MRI images of cancerous lesions;

[0079] S7. Use three-dimensional MRI images of cancerous lesions to train a lesion segmentation model to obtain a trained lesion segmentation model;

[0080] S8. Acquire the MRI image to be predicted, input the MRI image to be predicted into the trained cancer diagnosis model, output the MRI image to be predicted as cancer, input the MRI image to be predicted as cancer into the trained lesion segmentation model, and output the lesion segmentation result of the MRI image to be predicted as cancer.

[0081] This embodiment also acquired 3D MRI images of nasopharyngeal carcinoma patients from three hospital sites: Hospital A, Hospital B, and Hospital C. The 3D MRI images acquired from Hospital A were divided into a training set, a validation set, and an internal test set. The 3D MRI images acquired from Hospital B were divided into an external validation set 1, and the 3D MRI images acquired from Hospital C were divided into an external validation set 2. The nasopharyngeal carcinoma lesion segmentation method inspired by clinical decision-making processes proposed in this embodiment and the traditional deep classification method were used to diagnose cancer in the 3D MRI images from Hospitals A, B, and C, respectively. The cancer diagnosis results are shown in Table 1 and [Table data missing]. Figure 2 The figures in Table 1 represent the AUC (%), and the larger the value, the better the results of the corresponding method implementation;

[0082] Table 1. Diagnostic results of different methods in the nasopharyngeal carcinoma carcinoma diagnosis task.

[0083]

[0084] Furthermore, the nasopharyngeal carcinoma lesion segmentation method inspired by the clinical decision-making process proposed in this embodiment, as well as other recently disclosed nasopharyngeal carcinoma lesion segmentation methods, were used to segment lesions in the three-dimensional MRI images of hospitals A, B, and C. The lesion segmentation results are shown in Table 2 and [Table data missing]. Figure 3 The figures in Table 2 represent the DSC index (%), and the larger the value, the better the results of the corresponding method implementation;

[0085] Table 2. Lesion segmentation results of different methods in the cancer lesion segmentation task.

[0086]

[0087] The following conclusions can be drawn from Tables 1 and 2:

[0088] (1) Compared with traditional deep classification networks and traditional deep segmentation networks, the nasopharyngeal carcinoma lesion segmentation method inspired by clinical decision-making process proposed in this embodiment has achieved excellent performance in both the diagnosis and segmentation stages, verifying that the present invention can automatically complete high-precision diagnosis and segmentation.

[0089] (2) Compared with the performance at the primary lesion, all methods showed varying degrees of decline in diagnostic and segmentation performance at the lymph node, which is due to the more scattered distribution and blurred edges of the lymph nodes. However, the nasopharyngeal carcinoma lesion segmentation method inspired by the clinical decision-making process proposed in the example still achieved 96.4% AUC and 82.3% DSC, demonstrating the superiority of the carcinoma diagnosis model and lesion segmentation model.

[0090] (3) Compared with the deeper and more complex network structures used by DenseNet, GhostNet, NPC-Seg, AlignSeg and NPCNet, this invention achieves accurate results using only a lightweight ResNet-14 backbone network, verifying the effectiveness of the cancer diagnosis model and lesion segmentation model.

[0091] Example 2

[0092] See Figure 1 and Figure 4 The cancer diagnosis model described in S3 consists of a feature extraction submodule, a convolutional layer containing a 1x1 convolutional kernel, a sigmoid function activation layer, and a cancer probability diagnosis submodule connected in sequence. The feature extraction submodule is composed of a residual network without fully connected layers, and the cancer probability diagnosis submodule is composed of a parallel distribution alignment layer and an adaptive pooling layer.

[0093] See Figure 4 The specific processing procedure for inputting 3D MRI images into the cancer diagnosis model is as follows:

[0094] S31. Let the height of the 3D MRI image be H and the width be W. Input the 3D MRI image into the feature extraction submodule of the cancer diagnosis model. Use the feature extraction submodule to extract the depth features of the 3D MRI image and output the first depth feature map of the 3D MRI image. Perform a 1x1 convolution operation on the first depth feature map using a convolutional layer to output the second depth feature map. Perform a sigmoid transformation on the second depth feature map through a sigmoid function activation layer to transform the second depth feature map into a saliency map M, satisfying... To predict the probability of cancer in different regions of a 3D MRI image;

[0095] S32. An expert prior distribution alignment layer is used to align the saliency map, and an adaptive pooling layer is used to adaptively select k important pixels in the saliency map. The mean of these k important pixels is calculated, and the mean value is recorded as the cancer probability corresponding to the 3D MRI image, i.e., the image-level label prediction result of the 3D MRI image.

[0096] In step S32, the specific process of the distribution alignment layer performing expert prior distribution alignment on the saliency map is as follows:

[0097] S321. Identify the saliency map corresponding to the three-dimensional MRI images of batch n;

[0098] In step S321, for a batch of three-dimensional MRI images with batch number n... Repeat step S31 to obtain a batch of saliency maps.

[0099] S322. Calculate the probability mass distribution p of the saliency map described in S321. M p M The calculation expression is:

[0100]

[0101] Among them, M (i) Let represent the saliency map corresponding to the nth batch of 3D MRI images, and let ψ(.) regularization operation represent scaling the L1 norm of the entire matrix to 1;

[0102] S322. Determine the expert prior distribution of nasopharyngeal carcinoma in anatomical cross-sections. E The expert prior distribution p of nasopharyngeal carcinoma on anatomical cross sections. E Quantized to a Gaussian mixture distribution, p E The quantization expression is:

[0103]

[0104] in, Let π represent the i-th Gaussian distribution.i Represents the component weights and ∑π i =1, c represents the number of components, and the mean μ i The covariance Σ is determined by the anatomical location of the high-incidence sites of nasopharyngeal carcinoma. i This indicates locational perturbations caused by individual differences. According to the "Guidelines for Integrated Diagnosis and Treatment of Cancer in China", primary nasopharyngeal carcinoma lesions are most commonly found in the roof, posterior roof, and bilateral pharyngeal recesses of the nasopharynx. Lymphatic spread mainly occurs in the bilateral retropharyngeal and cervical lymph nodes. Therefore, the primary lesion site is defined as a four-component mixed distribution, and the lymph node site is defined as a two-component mixed distribution. By checking the location of high-incidence sites on the cross-section and fine-tuning on the validation set data, the parameters of each distribution can be obtained.

[0105] S323. Make the probability mass distribution p of the saliency graph... M Compared with expert prior distribution p E The distribution distance between them is the smallest;

[0106] In step S323, common distribution distance metrics include KL divergence, JS divergence, and optimal transmission distance. This step uses the optimal transmission distance, which is calculated as follows:

[0107]

[0108] Make T1 = vec(p) M ), T T 1 = vec(p E ), T≥0,

[0109] in, Let T denote the alignment loss, C denote the transfer matrix, and <·,·> denote the matrix dot product.

[0110] For any saliency map M of a three-dimensional MRI image obtained through step S31, calculate the size of the most important region and denote the number of important pixels as k. The formula for calculating k is:

[0111] k = card({M i,j :M i,j ≥δmax(M i,j )}),

[0112] Among them, M i,j Let represent the (i,j)th pixel of M, and δ∈[0,1] represent hyperparameters;

[0113] The average value of each pixel is recorded as the cancer probability corresponding to the 3D MRI image, which is the image-level label prediction result of the 3D MRI image. The calculation expression is as follows:

[0114]

[0115] S33. Determine the image-level label prediction results Does it exceed the diagnostic threshold? If so, then... The value is denoted as 1, indicating the image-level label prediction result. The corresponding 3D MRI image is a 3D MRI image of the cancerous lesion; otherwise, the 3D MRI image of the cancerous lesion will be output. Recorded as 0, this indicates the image-level label prediction result. The corresponding 3D MRI image is a normal 3D MRI image.

[0116] Example 3

[0117] See Figure 1 and Figure 5 The lesion segmentation model described in S6 consists of a connected fully convolutional encoder and decoder. The fully convolutional encoder is composed of a residual neural network without fully connected layers, and the decoder has parallel first and second convolutional branches. The specific processing procedure of the lesion segmentation model on the input three-dimensional MRI image of cancer is as follows:

[0118] S61. Multi-scale depth feature extraction from 3D MRI images of cancerous lesions using a fully convolutional encoder. Extract, where the k-th layer features F k The spatial dimension is The residual neural network here uses the same residual network as the feature extraction submodule, and uses the network parameters trained in step S31 as initialization; it outputs the multi-scale depth feature information of the three-dimensional MRI image of the cancer to the decoder;

[0119] S62. The decoder decodes the multi-scale depth feature information in a progressively refined manner, obtaining the pixel-level label prediction result P of the multi-scale depth feature information. k ;

[0120] In step S62, the specific decoding process is as follows:

[0121] S621. Transfer the k-th layer features F k The distributed input is fed into the first and second convolutional branches of the decoder. Both the first and second convolutional branches employ a network structure consisting of two cascaded 3x3 convolutional layers and one 1x1 convolutional layer. The first convolutional branch outputs a single-channel prediction result. The second convolutional branch outputs a confidence score map after a sigmoid transform. Confidence Score Plot C k Each pixel in the capture prediction result O k The prediction accuracy of the corresponding pixel; during the training phase, given the prediction result O k and pixel-level tag G kAmong them, G k Let G be the k-th element, and calculate the confidence score plot C. k Mean error loss The calculation expression is:

[0122]

[0123] Where σ(·) represents the sigmoid transform, G represents the Hadamard product. k Sampling from G 2 k The result is doubled;

[0124] S622. The predicted result O k Confidence Score Plot C k The Hadamard product and the prediction result U of the previous layer k+1 By fusion, the fused feature O' of the k-th layer is obtained. k O' k The calculation expression is:

[0125]

[0126] Where u(·) represents a 2x upsampling operation, U K+1 Represents a matrix consisting entirely of zeros;

[0127] S623. Fusion Feature O' k After cross-scale decoding by S622, two decoding results are output, one of which is the fused feature O' k After a 2x upsampling operation, the output U is used for the next layer fusion operation. k Another decoding result is the fusion feature O' k After sigmoid transformation, the segmentation prediction probability of this layer is output, and the segmentation prediction probability is denoted as the pixel-level label prediction result P. k P k The calculation expression is:

[0128]

[0129] S63. Determine the pixel-level label prediction result P k If the value is greater than the segmentation threshold, the pixel-level label prediction result is recorded as 1, indicating that the pixel in the 3D MRI image corresponding to the pixel-level label prediction result is a cancerous pixel; otherwise, the pixel-level label prediction result is recorded as 0, indicating that the pixel in the 3D MRI image corresponding to the pixel-level label prediction result is a normal pixel.

[0130] S64. Based on the pixel-level label prediction result P from step S63k The judgment result outputs the lesion segmentation result of the three-dimensional MRI image of cancer.

[0131] Example 4

[0132] When training the cancer diagnosis model, the training process uses the first loss function. As the total loss function, The expression is:

[0133]

[0134] in, Let λ represent the cross-entropy and λ represent the non-negative hyperparameter. Indicates alignment loss;

[0135] When training the lesion segmentation model, the training process uses the second loss function. and the third loss function As the total loss function, the second loss function The expression is:

[0136]

[0137] in, This represents the loss due to the mean square error.

[0138] Third loss function The expression is:

[0139]

[0140] in, G represents the cross-entropy. k Sampling from G 2 k The amount obtained is doubled.

[0141] Both the cancer diagnosis model and the lesion segmentation model used ResNet-18 without the last residual module as the backbone network. The size of all three-dimensional MRI images was scaled down to 512x512. λ=0.1, δ=0.45, the batch size of the cancer diagnosis model was 36, and the learning rate was 0.0001. The batch size of the lesion segmentation model was 4, and the learning rate was 0.001. An AMS-Grad optimizer with β1=0.9 and β2=0.999 was used. Gradient optimization was performed on the trained cancer diagnosis model and the trained lesion segmentation model using machine learning libraries such as PyTorch. The backpropagation process was completed to update the network parameters of the trained cancer diagnosis model and the trained lesion segmentation model, respectively.

[0142] Example 5

[0143] See Figure 6This embodiment also proposes a nasopharyngeal carcinoma lesion segmentation system inspired by clinical decision-making processes, the system comprising:

[0144] The acquisition module is used to acquire a certain number of three-dimensional MRI images of nasopharyngeal carcinoma patients to form a three-dimensional MRI image dataset.

[0145] The preprocessing module is used to preprocess the three-dimensional MRI images in the three-dimensional MRI image dataset;

[0146] The cancer diagnosis model construction module is used to construct a cancer diagnosis model based on three-dimensional MRI images for diagnosing cancer.

[0147] The cancer diagnosis model training module is used to divide the 3D MRI image dataset into a training set and a test set. The training set is used to train the cancer diagnosis model, and the test set is used to test the effectiveness of the cancer diagnosis model, so as to obtain a trained cancer diagnosis model.

[0148] The output module is used to input the three-dimensional MRI images from the acquisition module into the trained cancer diagnosis model and output the three-dimensional MRI images of the cancer.

[0149] The lesion segmentation model construction module is used to construct a lesion segmentation model for multi-scale segmentation of three-dimensional MRI images of cancerous lesions.

[0150] The lesion segmentation model training module is used to train a lesion segmentation model using three-dimensional MRI images of cancerous lesions, and obtain a trained lesion segmentation model.

[0151] The prediction module is used to acquire MRI images to be predicted, input the MRI images to be predicted into a trained cancer diagnosis model, and output the MRI images of cancer to be predicted. It also inputs the MRI images of cancer to be predicted into a trained lesion segmentation model and outputs the lesion segmentation results of the MRI images of cancer to be predicted.

[0152] In this embodiment, a deep learning framework is first established, from cancer diagnosis model to lesion segmentation model. First, the diagnostic model is used to diagnose pre-processed 3D MRI images, identifying cancerous 3D MRI images. The diagnostic process simulates a clinician's search for suspicious lesions starting from common sites of nasopharyngeal carcinoma and drawing a final conclusion from a local suspicious area. Then, the lesion segmentation model is used to perform multi-scale segmentation of the cancerous 3D MRI images, progressively refining the lesion segmentation results. The segmentation process simulates a clinician's multi-scale segmentation process after selecting different fields of view on the segmented object on a slice. The entire process from diagnosis to segmentation of the 3D MRI images is fully automated, without relying on prior diagnosis by a radiation oncologist, effectively improving the efficiency of nasopharyngeal carcinoma lesion segmentation and ensuring the interpretability of the nasopharyngeal carcinoma lesion results.

[0153] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A clinical decision process inspired nasopharyngeal carcinoma lesion segmentation method, characterized in that, Includes the following steps: S1. Obtain a certain number of three-dimensional MRI images of nasopharyngeal carcinoma patients and form a three-dimensional MRI image dataset; S2. Preprocess the three-dimensional MRI images in the three-dimensional MRI image dataset; S3. Construct a cancer diagnosis model for diagnosing cancer based on three-dimensional MRI images; the cancer diagnosis model consists of a feature extraction submodule, a convolutional layer containing a 1x1 convolutional kernel, a sigmoid function activation layer, and a cancer probability diagnosis submodule connected in sequence. The feature extraction submodule is composed of a residual network without fully connected layers, and the cancer probability diagnosis submodule is composed of parallel distribution alignment layers and adaptive pooling layers; the processing procedure of the cancer diagnosis model for three-dimensional MRI images is as follows: S31. Let the height of the three-dimensional MRI image be... Width The feature extraction submodule of the cancer diagnosis model inputs the three-dimensional MRI image into the feature extraction submodule, extracts the depth features of the three-dimensional MRI image, outputs the first depth feature map of the three-dimensional MRI image, performs a 1x1 convolution operation on the first depth feature map using a convolutional layer, outputs the second depth feature map, and performs a sigmoid transformation on the second depth feature map through a sigmoid function activation layer, transforming the second depth feature map into a saliency map. S32. Perform expert prior distribution alignment on the saliency map using a distribution alignment layer, and then utilize an adaptive pooling layer. Adaptive selection of saliency map Calculate 1 important pixel point The mean value of each important pixel is used to represent the cancer probability corresponding to the 3D MRI image, i.e., the image-level label prediction result of the 3D MRI image. ; The specific process of expert prior distribution alignment of the saliency map in the S32 distribution alignment layer is as follows: S321. Identify the saliency map corresponding to the three-dimensional MRI images of batch n; S322. Compute a probability mass distribution for the saliency map of S321 , The computational expression for this is: wherein, S n represents a saliency map corresponding to the n-th batch of three-dimensional MRI images, The regularization operation represents scaling the L1 norm of the entire matrix to 1. S322. Determine the expert prior distribution of nasopharyngeal carcinoma in anatomical cross sections. The expert a priori distribution of nasopharyngeal carcinoma on anatomical cross sections Quantized into a Gaussian mixture distribution, The quantization expression is: in, Indicates the first A Gaussian distribution, Indicates component weights and , Indicates the number of components, mean The covariance is determined by the anatomical location of the high-incidence sites of nasopharyngeal carcinoma. This indicates positional perturbations caused by individual differences; S323. Massaging the map probability mass distribution with the expert prior distribution to minimize the distribution distance between them S33. Determine the image-level label prediction results Does it exceed the diagnostic threshold? If so, then... The value is 1, indicating the image-level label prediction result. The corresponding 3D MRI image is a 3D MRI image of the cancerous lesion; otherwise, the 3D MRI image of the cancerous lesion will be output. Recorded as 0, this indicates the image-level label prediction result. The corresponding 3D MRI image is a normal 3D MRI image; S4. Divide the 3D MRI image dataset into a training set and a test set. Use the training set to train the cancer diagnosis model and use the test set to test the effectiveness of the cancer diagnosis model to obtain a well-trained cancer diagnosis model. S5. Input the preprocessed 3D MRI images from S2 into the trained cancer diagnosis model and output the 3D MRI images of the cancer. S6. Construct a lesion segmentation model for multi-scale segmentation of three-dimensional MRI images of cancerous lesions; S7. Use three-dimensional MRI images of cancerous lesions to train a lesion segmentation model to obtain a trained lesion segmentation model; S8. Acquire the MRI image to be predicted, input the MRI image to be predicted into the trained cancer diagnosis model, output the MRI image to be predicted as cancer, input the MRI image to be predicted as cancer into the trained lesion segmentation model, and output the lesion segmentation result of the MRI image to be predicted as cancer.

2. The clinicai decision support process enabled nasopharyngeal carcinoma lesion segmentation method according to claim 1, characterized in that, The three-dimensional MRI images of the nasopharyngeal carcinoma patient described in step S1 include T1-weighted or T1-weighted contrast-enhanced images of the head and T2-weighted images of the neck.

3. The clinicai decision support process enabled nasopharyngeal carcinoma lesion segmentation method according to claim 1, wherein, Step S2 involves preprocessing the three-dimensional MRI images in the three-dimensional MRI image dataset by using the maximum and minimum value normalization operation to scale the pixel values ​​of the three-dimensional MRI images to the range of [0,1].

4. The clinicai decision support process enabled nasopharyngeal carcinoma lesion segmentation method according to claim 3, characterized in that, Any 3D MRI image in the S4 training set is matched with an image-level label. and pixel-level tags ,when This indicates that the 3D MRI image is a normal 3D MRI image; when When the image is in a certain state, it indicates that the 3D MRI image belongs to a cancerous 3D MRI image; when... When the pixel in the 3D MRI image is normal, it indicates that the pixel is normal. When the value is 0, it indicates that the pixel in the 3D MRI image is a cancerous pixel.

5. The clinic-decision-process steered nasopharyngeal carcinoma lesion segmentation method according to claim 4, characterized in that, The lesion segmentation model described in S6 consists of a connected fully convolutional encoder and decoder. The specific processing procedure of the lesion segmentation model on the input three-dimensional MRI image of cancerous lesions is as follows: S61. Use a fully convolutional encoder to extract multi-scale depth features from the three-dimensional MRI images of cancerous lesions, and output the multi-scale depth feature information of the three-dimensional MRI images of cancerous lesions to the decoder. S62. The decoder decodes the multi-scale depth feature information in a step-by-step refinement manner to obtain a pixel-level label prediction result of the multi-scale depth feature information ; S63. Determine the pixel-level label prediction result If the value is greater than the segmentation threshold, the pixel-level label prediction result is recorded as 1, indicating that the pixel in the 3D MRI image corresponding to the pixel-level label prediction result is a cancerous pixel; otherwise, the pixel-level label prediction result is recorded as 0, indicating that the pixel in the 3D MRI image corresponding to the pixel-level label prediction result is a normal pixel. S64. Based on the pixel-level label prediction results of step S63 The judgment result outputs the lesion segmentation result of the three-dimensional MRI image of cancer.

6. The nasopharyngeal carcinoma lesion segmentation method inspired by clinical decision-making processes according to claim 5, characterized in that, When training the cancer diagnosis model, the training process uses the first loss function. As the total loss function, The expression is: wherein, denotes cross-entropy, denotes non-negative hyperparameters, denotes alignment loss; In training the lesion segmentation model, the training process uses the second loss function and the third loss function as a total loss function, the expression of the second loss function is: wherein, denotes the mean squared error loss; Third loss function The expression is: wherein, denotes the cross-entropy, by down-sampling times.

7. A nasopharyngeal carcinoma lesion segmentation system inspired by clinical decision-making processes, said system being implemented based on the method of claim 1, characterized in that, The system includes: The acquisition module is used to acquire a certain number of three-dimensional MRI images of nasopharyngeal carcinoma patients to form a three-dimensional MRI image dataset. The preprocessing module is used to preprocess the three-dimensional MRI images in the three-dimensional MRI image dataset; The cancer diagnosis model construction module is used to construct a cancer diagnosis model based on three-dimensional MRI images for diagnosing cancer. The cancer diagnosis model training module is used to divide the 3D MRI image dataset into a training set and a test set. The training set is used to train the cancer diagnosis model, and the test set is used to test the effectiveness of the cancer diagnosis model, so as to obtain a trained cancer diagnosis model. The output module is used to input the three-dimensional MRI images from the acquisition module into the trained cancer diagnosis model and output the three-dimensional MRI images of the cancer. The lesion segmentation model construction module is used to construct a lesion segmentation model for multi-scale segmentation of three-dimensional MRI images of cancerous lesions. The lesion segmentation model training module is used to train a lesion segmentation model using three-dimensional MRI images of cancerous lesions, and obtain a trained lesion segmentation model. The prediction module is used to acquire MRI images to be predicted, input the MRI images to be predicted into a trained cancer diagnosis model, and output the MRI images of cancer to be predicted. It also inputs the MRI images of cancer to be predicted into a trained lesion segmentation model and outputs the lesion segmentation results of the MRI images of cancer to be predicted.

Images