Cervical cancer MRI image tumor segmentation method based on nnU-Net network

Abstract

The application relates to the technical field of image processing, in particular to a cervical cancer MRI image tumor segmentation method based on an nnU-Net network, which comprises the following steps: collecting cervical cancer MRI images of cervical cancer patients, removing invalid backgrounds of the cervical cancer MRI images, performing resampling processing and edge slice on the cervical cancer MRI images to output optimized images; training an initial model according to image features to output an nnU-Net model, automatically segmenting the cervical cancer MRI images by using the nnU-Net model to output tumor region images and non-tumor region images; determining whether the accuracy of cervical cancer MRI image segmentation meets the requirements based on the signal-to-noise ratio of the optimized images; and adjusting the edge slice overlap rate; if the effectiveness of the nnU-Net model training does not meet the requirements, the batch training size of the preprocessed data is adjusted. The application improves the accuracy of cervical cancer MRI image segmentation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a cervical cancer MRI image tumor segmentation method based on an nnU-Net network. BACKGROUND

[0002] In the prior art, cervical cancer (CC) is one of the main causes of cancer-related deaths in women worldwide. Persistent infection of high-risk HPV is a clear pathogenic factor for the occurrence and development of cervical cancer. With the rapid development of modern medicine, cervical cancer has become the first cancer that can be jointly prevented and controlled by HPV vaccine and systematic screening, but even so, it is still the main cause of cancer death among women in developing countries. The problems involved in early screening and HPV vaccination, such as high-risk groups, high-risk areas, high-risk occupations, and high-risk factors, need to be addressed urgently. Magnetic resonance imaging (MRI) plays an indispensable key role in the diagnosis and treatment of cervical cancer. With its high resolution and multi-parameter imaging technology, MRI can accurately identify and distinguish cervical cancer lesions and assist in clinical staging.

[0003] Chinese Patent Publication No. CN102999917A discloses a cervical cancer image automatic segmentation method based on T2-weighted magnetic resonance imaging (T2-MRI) and diffusion-weighted magnetic resonance imaging (DW-MRI), comprising: step 1: using a nonlinear registration method to register the DW-MR image to the T2-MR image, and classifying the registered DW-MR image; step 2: using a nonlinear anisotropic diffusion filtering technique to filter the T2-MR image, segmenting the bladder and rectum, and using the segmentation results of the bladder and rectum to segment the region of interest; step 3: using a joint maximum a posteriori probability (CMAP) method to accurately segment the tumor in the region of interest of the T2-MR image and the DW-MR image.

[0004] As can be seen, the prior art has the following problems: if the parameters are not properly set, the image may be excessively smoothed, resulting in loss of tumor boundary and other detailed information, affecting the accuracy of subsequent segmentation. SUMMARY

[0005] To this end, the present application provides a cervical cancer MRI image tumor segmentation method based on an nnU-Net network to overcome the problem in the prior art that if the parameters are not properly set, the image may be excessively smoothed, resulting in loss of tumor boundary and other detailed information, affecting the accuracy of subsequent segmentation.

[0006] To achieve the above-mentioned purpose, the present application provides a cervical cancer MRI image tumor segmentation method based on an nnU-Net network, comprising:

[0007] Collecting a cervical cancer MRI image of a cervical cancer patient, removing invalid background from the cervical cancer MRI image, performing resampling processing and edge slice to output an optimized image;

[0008] Extracting image features from the optimized image, training an initial model according to the image features to output an nnU-Net model, updating the nnU-Net model by calculating the gradient of each parameter compared to the loss function, and using the nnU-Net model to automatically segment the cervical cancer MRI image to output a tumor region image and a non-tumor region image;

[0009] Determining whether the accuracy of the cervical cancer MRI image segmentation meets the requirements based on the signal-to-noise ratio of the optimized image;

[0010] If the accuracy of the cervical cancer MRI image segmentation does not meet the requirements, adjusting the edge slice overlap rate, or determining whether the effectiveness of the nnU-Net model training meets the requirements based on the recall rate of the nnU-Net model;

[0011] If the effectiveness of the nnU-Net model training does not meet the requirements, adjusting the batch training size of the preprocessed data, or adjusting the learning rate warm restart period of the nnU-Net model based on the gradient sign change rate of adjacent training batches.

[0012] Further, compare the signal-to-noise ratio of the optimized image with a preset second signal-to-noise ratio;

[0013] If the signal-to-noise ratio of the optimized image is less than or equal to the second signal-to-noise ratio, it is determined that the accuracy of the cervical cancer MRI image segmentation does not meet the requirements.

[0014] Further, compare the signal-to-noise ratio of the optimized image with the preset first signal-to-noise ratio and the preset second signal-to-noise ratio, respectively;

[0015] If the signal-to-noise ratio of the optimized image is greater than the preset first signal-to-noise ratio and less than or equal to the preset second signal-to-noise ratio, it is preliminarily determined that the effectiveness of the nnU-Net model training does not meet the requirements, and whether the effectiveness of the nnU-Net model training meets the requirements is determined according to the recall rate of the nnU-Net model.

[0016] Further, compare the signal-to-noise ratio of the optimized image with the preset first signal-to-noise ratio;

[0017] If the signal-to-noise ratio of the optimized image is less than the preset first signal-to-noise ratio, increase the edge slice overlap rate.

[0018] Further, the increase in the edge slice overlap rate is determined by the difference between the signal-to-noise ratio of the optimized image and the preset first signal-to-noise ratio.

[0019] Further, the recall rate of the nnU-Net model is compared with a preset first recall rate and a preset second recall rate respectively.

[0020] If the recall rate of the nnU-Net model is greater than the preset second recall rate, it is determined that the effectiveness of the nnU-Net model training meets the requirements.

[0021] If the recall rate of the nnU-Net model is less than or equal to the preset first recall rate, it is preliminarily determined that the stability of the loss function does not meet the requirements, and the stability of the loss function is determined according to the gradient sign change rate of adjacent training batches.

[0022] If the recall rate of the nnU-Net model is greater than the preset first recall rate and less than or equal to the preset second recall rate, the batch training size of the preprocessed data is reduced.

[0023] Further, the reduction range of the batch training size of the preprocessed data is determined by the difference between the recall rate of the nnU-Net model and the preset first recall rate.

[0024] Further, the gradient sign change rate of adjacent training batches is compared with a preset change rate.

[0025] If the gradient sign change rate of adjacent training batches is greater than the preset change rate, it is determined that the stability of the loss function does not meet the requirements, and the learning rate warm restart period of the nnU-Net model is reduced.

[0026] Further, the gradient sign change rate of adjacent training batches is the ratio of the number of elements with changed gradient signs in adjacent two batches to the total number of parameter elements.

[0027] Further, the reduction range of the learning rate warm restart period of the nnU-Net model is determined by the difference between the gradient sign change rate of adjacent training batches and the preset change rate.

[0028] Compared with the prior art, the method has the beneficial effects that the method adjusts the edge slice overlap rate according to the signal-to-noise ratio of the optimized image, because the magnetic field strength gradually weakens and the uniformity increases in the edge area of the MRI magnet, when the body part of the examinee is located in the edge magnetic field area, due to the changes of the magnetic field strength and the uniformity, the signal acquisition of the part of the tissue is inaccurate, and the image quality is poor, by increasing the edge slice overlap rate, the algorithm can refer to more context information when processing the edge area, which helps to more accurately identify and segment the tumor boundary, reduces the segmentation error caused by the magnetic field non-uniformity, adjusts the batch training size of the preprocessed data according to the recall rate of the nnU-Net model, because the nnU-Net model has a large step size, the model is down-sampled too fast, a large amount of image detail information is lost, and some subtle features of the tumor cannot be extracted, by reducing the batch training size of the preprocessed data, the amount of data participating in gradient calculation in each training step is small, the model updates more frequently, which helps to more carefully estimate the gradient when the model is down-sampled too fast and a large amount of image detail information is lost, and capture subtle features that are easily ignored, and the learning rate warm restart period of the nnU-Net model is adjusted according to the gradient sign change rate of adjacent training batches, because in a long training process, some parameters of the model may degenerate, some parameters may become too large or too small, causing the gradient to disappear or explode, and then affecting the stability of the loss function, by reducing the learning rate warm restart period of the nnU-Net model, the learning rate can be restored to a larger value in time, so that the model can more actively adjust the parameters and adapt to the changes of the data distribution, which helps to overcome the parameter degradation problem.

[0029] Further, the method adjusts the edge slice overlap rate by setting the preset first signal-to-noise ratio and the preset second signal-to-noise ratio, because the changes of the magnetic field strength and the uniformity will cause the signal acquisition of the part of the tissue to be inaccurate and the image quality to be poor, by increasing the edge slice overlap rate, the algorithm can refer to more context information when processing the edge area, which helps to more accurately identify and segment the tumor boundary, reduces the segmentation error caused by the magnetic field non-uniformity, and improves the accuracy of cervical cancer MRI image segmentation.

[0030] Further, the method adjusts the batch training size of the preprocessed data by setting a preset first recall rate and a preset second recall rate. Since the nnU-Net model has a large step size, the model may be down-sampled too fast, a large amount of image detail information is lost, and some subtle features of the tumor cannot be extracted. By reducing the batch training size of the preprocessed data, the amount of data participating in gradient calculation in each training step is reduced, and the model is updated more frequently, which helps to estimate the gradient more carefully when the model is down-sampled too fast and a large amount of image detail information is lost, and capture subtle features that are easily overlooked, thereby further improving the accuracy of cervical cancer MRI image segmentation.

[0031] Further, the method adjusts the learning rate warm restart cycle of the nnU-Net model by setting a preset change rate. During a long training process, model parameters may degenerate, some parameters may become too large or too small, resulting in gradient vanishing or explosion, and thus affecting the stability of the loss function. By reducing the learning rate warm restart cycle of the nnU-Net model, the learning rate can be restored to a larger value in a timely manner, so that the model can more actively adjust the parameters and adapt to changes in data distribution, which helps to overcome the parameter degradation problem and further improves the accuracy of cervical cancer MRI image segmentation. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 FIG. 1 is a flowchart of a cervical cancer MRI image tumor segmentation method based on a nnU-Net network according to an embodiment of the present application;

[0033] Figure 2 FIG. 2 is a logic flowchart of an edge slice overlap rate adjustment process of a cervical cancer MRI image tumor segmentation method based on a nnU-Net network according to an embodiment of the present application;

[0034] Figure 3 FIG. 3 is a logic flowchart of a batch training size adjustment process of preprocessed data of a cervical cancer MRI image tumor segmentation method based on a nnU-Net network according to an embodiment of the present application;

[0035] Figure 4 FIG. 4 is a logic flowchart of a learning rate warm restart cycle adjustment process of a nnU-Net model of a cervical cancer MRI image tumor segmentation method based on a nnU-Net network according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] In order to make the objects and advantages of the present application clearer, the following further describes the present application with reference to the embodiments; it should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0037] Preferred embodiments of the present application will be described herein below with reference to the accompanying drawings. Those skilled in the art will appreciate that the embodiments are only used to explain the technical principles of the present application, and are not intended to limit the scope of protection of the present application.

[0038] Please refer to Figure 1 , Figure 2 , Figure 3 and Figure 4 , which are respectively the overall flowchart of the cervical cancer MRI image tumor segmentation method based on the nnU-Net network of the embodiment of the present application, the logic flowchart of the adjustment process of the edge slice overlap rate, the logic flowchart of the adjustment process of the batch training size of the preprocessed data, and the logic flowchart of the adjustment process of the learning rate warm restart period of the nnU-Net model. The cervical cancer MRI image tumor segmentation method based on the nnU-Net network of the present application comprises:

[0039] Step S1, collecting the cervical cancer MRI image of the cervical cancer patient, removing the invalid background of the cervical cancer MRI image, performing resampling processing and edge slice to output an optimized image;

[0040] Step S2, performing image feature extraction on the optimized image, training an initial model according to the image features to output an nnU-Net model, updating the nnU-Net model by calculating the gradient of each parameter compared to the loss function, and using the nnU-Net model to automatically segment the cervical cancer MRI image to output a tumor region image and a non-tumor region image;

[0041] Step S3, determining whether the accuracy of the cervical cancer MRI image segmentation meets the requirements based on the signal-to-noise ratio of the optimized image;

[0042] Step S4, if the accuracy of the cervical cancer MRI image segmentation does not meet the requirements, adjusting the edge slice overlap rate, or determining whether the effectiveness of the nnU-Net model training meets the requirements based on the recall rate of the nnU-Net model;

[0043] Step S5, if the effectiveness of the nnU-Net model training does not meet the requirements, adjusting the batch training size of the preprocessed data, or adjusting the learning rate warm restart period of the nnU-Net model based on the gradient sign change rate of adjacent training batches;

[0044] Specifically, the optimized image includes the cervical cancer MRI image after removing the invalid background, the cervical cancer MRI image after resampling processing, and the cervical cancer MRI image after edge slice.

[0045] Specifically, the loss function includes cross-entropy loss and dice coefficient similarity loss.

[0046] Specifically, the formula of the loss function is:

[0047]

[0048]

[0049]

[0050] Specifically, the signal-to-noise ratio of the optimized image is the ratio of the power of the tumor tissue organ information signal to the power of the random noise in the image.

[0051] Specifically, the slice overlap rate of the tumor edge in the segmentation result is the ratio of the area of the overlapping part between the slices of the tumor edge in the segmentation result and the slices of the real tumor edge to the total area of the two.

[0052] Specifically, the recall rate of the nnU-Net model is the ratio of the number of tumor region pixels correctly predicted by the model to the number of real tumor region pixels.

[0053] Specifically, the formula of the recall rate is:

[0054]

[0055] Specifically, the batch training size of the preprocessed data is the number of samples used to update the model parameters in each iteration.

[0056] Specifically, the gradient sign change rate of adjacent training batches is the ratio of the number of elements with changed gradient signs in the two adjacent training batches to the total number of gradient elements.

[0057] Specifically, the gradient sign usually refers to the representation of the direction of the change of the pixel value in the image.

[0058] Specifically, the elements refer to the individual components in the gradient vector.

[0059] Specifically, the learning rate warm restart period of the nnU-Net model is the training rounds in which the learning rate is periodically changed in the form of a cosine function during training, and in each period, the learning rate gradually decreases from a larger value to a smaller value, and then returns to a larger value.

[0060] Specifically, the nnU-Net model includes 2D U-Net and 3D U-Net.

[0061] The workflow of U-Net covers several key steps, which are as follows:

[0062] (1) Convolution: Similar to the original U-Net, nnU-Net performs two convolution operations on the input image in the encoder part, each followed by a nonlinear layer via leaky ReLU and a normalization layer of instance normalization. This configuration is beneficial to expand the number of channels while maintaining the image pixel resolution, achieving abstract feature extraction;

[0063] (2) Down-sampling: Down-sampling is achieved by stride convolution, each round of down-sampling halves the image size while doubling the number of channels. This operation is beneficial to remove invalid information, retain main features, and improve algorithm efficiency. Up-sampling is performed using transposed convolution to restore the image size to the initial scale. This process aims to restore the image details reduced by down-sampling and effectively integrate the information in the down-sampling process;

[0064] (3) Up-sampling: Up-sampling is performed using transposed convolution to restore the image size to the initial scale. The purpose of this operation is to restore the image details reduced by down-sampling and effectively integrate the information in the down-sampling process;

[0065] (4) Information fusion: nnU-Net achieves information fusion between the encoder and decoder through skip connections. This structure can effectively combine feature information of different resolutions and preserve the spatial details of the image;

[0066] (5) Output prediction: After up-sampling and information fusion, the feature map is restored to the original image size. Finally, a 1x1 convolution operation is performed for classification, and a softmax function is used for probabilistic processing to obtain the final segmentation prediction result.

[0067] Specifically, in the 2D U-Net configuration, the 3D image is segmented into 2D slices with a size of 384x384. Subsequently, the input image is mapped to 32 channels through the first convolution operation, followed by instance normalization and Leaky ReLU activation processes. As the network deepens, the spatial dimension gradually decreases, while the number of feature extraction channels in each downsampling layer increases step by step. After that, the spatial dimension decreases, and the number of feature extraction channels in each downsampling layer increases. The bottom layer feature dimension is 512x6x6, and the bottom layer feature is transmitted to the upsampling layer through the jump connection to return to the original dimension. Each upsampling stage includes transpose convolution, instance normalization, and LeakyReLU. Finally, the convolution reduces the number of channels to the number of segmentation categories, and the sigmoid activation is applied to generate the final segmentation map. In the bottom layer, the feature dimension is compressed to 512x6x6. The bottom layer feature then enters the upsampling layer and recovers to the original spatial dimension through the jump connection. Each upsampling stage includes transpose convolution, instance normalization, and Leaky ReLU activation. Finally, the number of channels is reduced to the number of segmentation categories through 1x1 convolution, and the sigmoid activation function is applied to generate the final segmentation result. This hierarchical architecture effectively combines global background information and local detailed features, achieving accurate segmentation of medical images.

[0068] For the nnU-Net with 3D U-Net configuration, the 3D image is divided into 24x256x256 pixel three-dimensional patches. Unlike the 2D U-Net configuration of nnU-Net, in each stage, 2D convolution and 2D transpose convolution are replaced by 3D convolution and 3D transpose convolution. During this process, instance normalization and Leaky ReLU activation functions are still used. Through this configuration, the model can better capture the structural information in three-dimensional space, further improving the accuracy of segmentation. The training loss combines the Dice coefficient similarity loss (L_DSC) and the cross-entropy loss (L_CE).

[0069] Each model was trained for a total of 1000 iterations. During training, the stochastic gradient descent method (SGD) was used for optimization, with an initial learning rate of 0.01 and the application of Nesterov momentum with a value of 0.99. To evaluate model performance and prevent overfitting, we used a five-fold cross-validation method to ensure the robustness, generalization ability, and reliability of the entire segmentation process. Finally, the segmentation model was independently tested using MRI images from 30 patients. After training, the segmentation results generated by the optimal model need to be post-processed to improve accuracy. The nnU-Net framework can automatically adjust all parameters related to magnetic resonance imaging.

[0070] In the implementation, the method adjusts the edge slice overlap rate according to the signal-to-noise ratio of the optimized image. Since the magnetic field strength gradually weakens and the uniformity increases in the edge region of the MRI magnet, when the body part of the subject is located in the edge magnetic field region, the signal acquisition of the part of the tissue is inaccurate due to the change of the magnetic field strength and uniformity, and the image quality is reduced. By increasing the edge slice overlap rate, the algorithm can refer to more context information when processing the edge region, which helps to more accurately identify and segment the tumor boundary, reduces the segmentation error caused by the magnetic field non-uniformity, adjusts the batch training size of the pre-processed data according to the recall rate of the nnU-Net model. Since the nnU-Net model has a large step size, the model down-samples too quickly, loses a lot of image detail information, and some subtle features of the tumor cannot be extracted. By reducing the batch training size of the pre-processed data, the amount of data participating in gradient calculation in each training step is smaller, and the model updates more frequently, which helps to estimate the gradient more carefully when the model down-samples too quickly and loses a lot of image detail information, and capture subtle features that are easily overlooked. Adjust the learning rate warm restart period of the nnU-Net model according to the gradient sign change rate of adjacent training batches. Since some degradation may occur in the model parameters during a long training process, some parameters may become too large or too small, causing the gradient to disappear or explode, thereby affecting the stability of the loss function. By reducing the learning rate warm restart period of the nnU-Net model, the learning rate can be restored to a larger value in a timely manner, allowing the model to more actively adjust the parameters and adapt to changes in data distribution, which helps to overcome the parameter degradation problem.

[0071] Specifically, determining the accuracy of the cervical cancer MRI image segmentation includes:

[0072] Comparing the signal-to-noise ratio of the optimized image with a preset second signal-to-noise ratio;

[0073] If the signal-to-noise ratio of the optimized image is less than or equal to the second signal-to-noise ratio, it is determined that the accuracy of the cervical cancer MRI image segmentation does not meet the requirements.

[0074] Specifically, determining the effectiveness of the nnU-Net model training includes:

[0075] Comparing the signal-to-noise ratio of the optimized image with the preset first signal-to-noise ratio and the preset second signal-to-noise ratio, respectively;

[0076] If the signal-to-noise ratio of the optimized image is greater than the preset first signal-to-noise ratio and less than or equal to the preset second signal-to-noise ratio, it is preliminarily determined that the effectiveness of the nnU-Net model training does not meet the requirements, and whether the effectiveness of the nnU-Net model training meets the requirements is determined according to the recall rate of the nnU-Net model.

[0077] It can be understood that the three intervals divided by the preset first signal-to-noise ratio and the preset second signal-to-noise ratio correspond to three situations respectively.

[0078] The first interval is that the signal-to-noise ratio of the optimized image is less than the preset first signal-to-noise ratio, and the corresponding situation is that due to the change of the magnetic field strength and uniformity, the signal acquisition of the part of the tissue is inaccurate, and the image quality is reduced.

[0079] The second interval is that the signal-to-noise ratio of the optimized image is greater than the preset first signal-to-noise ratio and less than the preset second signal-to-noise ratio, and the corresponding situation is that due to the too large step of the nnU-Net model, the model is down-sampled too fast, a large amount of image detail information is lost, and some subtle features of the tumor cannot be extracted.

[0080] The third interval is that the signal-to-noise ratio of the optimized image is greater than the preset second signal-to-noise ratio, and the corresponding situation is that the accuracy of the cervical cancer MRI image segmentation is determined to meet the requirements.

[0081] It can be understood that the preset signal-to-noise ratio can be set according to the actual working condition, and the preset signal-to-noise ratio is intended to ensure the accuracy and practicability of the test results. Alternatively, the preset signal-to-noise ratio is determined by a series of pre-experiments, cervical cancer MRI images with different signal-to-noise ratios are collected and processed, and then the nnU-Net model is used for tumor segmentation experiment, and the accuracy and recall rate of the segmentation results are analyzed to determine an optimal signal-to-noise ratio range. Exemplarily, the preset first signal-to-noise ratio is generally selected in the range of [10dB, 15dB], and the preset second signal-to-noise ratio is generally selected in the range of [16dB, 20dB].

[0082] Preferably, the preferred embodiment of the preset first signal-to-noise ratio is 13dB, and the preferred embodiment of the preset second signal-to-noise ratio is 18dB.

[0083] Specifically, adjusting the edge slice overlap rate comprises:

[0084] Comparing the signal-to-noise ratio of the optimized image with the preset first signal-to-noise ratio;

[0085] If the signal-to-noise ratio of the optimized image is less than the preset first signal-to-noise ratio, the edge slice overlap rate is increased.

[0086] Specifically, when the difference between the signal-to-noise ratio of the optimized image and the preset first signal-to-noise ratio is within 2dB, the edge slice overlap rate is increased to 1.2 times of the original, and when the difference between the signal-to-noise ratio of the optimized image and the preset first signal-to-noise ratio exceeds 2dB, the edge slice overlap rate is increased by 2% for each 1dB exceeding 1dB on the basis of being increased to 1.2 times of the original, for example, when the difference between the signal-to-noise ratio of the optimized image and the preset first signal-to-noise ratio is 4dB, and the current edge slice overlap rate is 25%, the increased edge slice overlap rate is 25*1.2+2*2=34%.

[0087] In implementation, the method provided by the application adjusts the edge slice overlap rate by the preset first signal-to-noise ratio and the preset second signal-to-noise ratio. Due to the change of the magnetic field intensity and uniformity, the signal acquisition of the part of the tissue is inaccurate, and the image quality is reduced. By increasing the edge slice overlap rate, the algorithm can refer to more context information when processing the edge area, which helps to more accurately identify and segment the tumor boundary, reduces the segmentation error caused by the magnetic field non-uniformity, and improves the accuracy of cervical cancer MRI image segmentation.

[0088] Specifically, the increase amplitude of the edge slice overlap rate is determined by the difference between the signal-to-noise ratio of the optimized image and the preset first signal-to-noise ratio.

[0089] Specifically, the batch training size of the preprocessed data is adjusted, including:

[0090] The recall rate of the nnU-Net model is compared with the preset first recall rate and the preset second recall rate respectively;

[0091] If the recall rate of the nnU-Net model is greater than the preset second recall rate, it is determined that the effectiveness of the nnU-Net model training meets the requirements;

[0092] If the recall rate of the nnU-Net model is less than or equal to the preset first recall rate, it is preliminarily determined that the stability of the loss function does not meet the requirements, and the stability of the loss function is determined according to the gradient sign change rate of the adjacent training batches;

[0093] If the recall rate of the nnU-Net model is greater than the preset first recall rate and less than or equal to the preset second recall rate, the batch training size of the preprocessed data is reduced.

[0094] It can be understood that the three intervals divided by the preset first recall rate and the preset second recall rate correspond to three situations respectively:

[0095] The first interval is that the recall rate of the nnU-Net model is less than or equal to a preset first recall rate. In this case, due to the long training process, some degradation may occur in the model parameters, some parameters may become too large or too small, causing gradient disappearance or explosion, and thus affecting the stability of the loss function.

[0096] The second interval is that the recall rate of the nnU-Net model is greater than the preset first recall rate and less than or equal to a preset second recall rate. In this case, due to the large step size of the nnU-Net model, the model may downsample too quickly, losing a large amount of image detail information, and thus some subtle features of the tumor cannot be extracted.

[0097] The third interval is that the recall rate of the nnU-Net model is greater than the preset second recall rate. In this case, it is determined that the nnU-Net model training is effective.

[0098] It can be understood that the preset recall rate can be set according to the model evaluation. The preset recall rate is intended to ensure the accuracy and practicality of the test results. Alternatively, the preset recall rate is determined by evaluating the performance of the model at different recall rates on the validation set, observing the segmentation effect of the model on the validation set, and selecting an optimal preset recall rate. For example, the preset first recall rate is generally selected in the range of [0.7, 0.8], and the preset second recall rate is generally selected in the range of [0.85, 0.95].

[0099] Preferably, the preferred embodiment of the preset first recall rate is 0.75, and the preferred embodiment of the preset second recall rate is 0.9.

[0100] Specifically, the reduction range of the batch training size of the preprocessed data is determined by the difference between the recall rate of the nnU-Net model and the preset first recall rate.

[0101] Specifically, when the difference between the recall rate of the nnU-Net model and the preset first recall rate is within 0.1, the batch training size of the preprocessed data is reduced to 0.9 times the original size. When the difference between the recall rate of the nnU-Net model and the preset first recall rate exceeds 0.1, the batch training size of the preprocessed data is reduced by 3 for each 0.05 exceeded, based on the reduction to 0.9 times the original size. For example, when the difference between the recall rate of the nnU-Net model and the preset first recall rate is 0.15, the batch training size of the preprocessed data is 32, and the reduced batch training size of the preprocessed data is 32x0.9-3x1≈26.

[0102] Specifically, when the reduced batch training size of the preprocessed data is a decimal, the decimal is automatically carried forward and the integer is retained.

[0103] In the implementation, the method disclosed by the application adjusts the batch training size of the preprocessed data by setting a preset first recall rate and a preset second recall rate. Due to the excessively large step size of the nnU-Net model, the model may be down-sampled too fast, a large amount of image detail information may be lost, and some subtle features of the tumor may not be extracted. By reducing the batch training size of the preprocessed data, the amount of data participating in gradient calculation in each training step can be reduced, and the model can be updated more frequently, which helps to estimate the gradient more carefully when the model is down-sampled too fast and a large amount of image detail information is lost, and capture subtle features that are easily overlooked, thereby further improving the accuracy of cervical cancer MRI image segmentation.

[0104] Specifically, the learning rate warm restart period of the nnU-Net model is adjusted, including:

[0105] Comparing the gradient sign change rate of adjacent training batches with a preset change rate;

[0106] If the gradient sign change rate of the adjacent training batches is greater than the preset change rate, it is determined that the stability of the loss function does not meet the requirements, and the learning rate warm restart period of the nnU-Net model is reduced.

[0107] Specifically, the gradient sign change rate of the adjacent training batches is the ratio of the number of elements whose gradient signs change in adjacent two batches to the total number of parameter elements.

[0108] It can be understood that the two intervals divided by the preset change rate correspond to two situations respectively.

[0109] The first interval is that the gradient sign change rate of adjacent training batches is less than the preset change rate, and the corresponding situation is that the stability of the loss function meets the requirements.

[0110] The second interval is that the gradient sign change rate of adjacent training batches is greater than or equal to the preset change rate, and the corresponding situation is that due to the degradation of the model parameters in a long training process, some parameters may become too large or too small, resulting in gradient disappearance or explosion, and thus affecting the stability of the loss function.

[0111] It can be understood that the preset change rate can be set according to the actual working condition. The preset change rate is intended to ensure the accuracy and practicability of the test results. Optionally, the preset change rate is calculated by calculating the gradient of the model parameters for each training batch during the training of the deep learning model, and the stability of the parameter update in the model training process can be monitored. For example, the preset change rate is generally selected in the range of [10%, 15%], and preferably, the preferred embodiment of the preset change rate is 13%.

[0112] Specifically, the reduction range of the learning rate warm restart cycle of the nnU-Net model is determined by the difference between the gradient sign change rate of adjacent training batches and the preset change rate.

[0113] Specifically, when the difference between the gradient sign change rate of adjacent training batches and the preset change rate is within 2%, the learning rate warm restart cycle of the nnU-Net model is reduced to 0.9 times of the original; when the difference between the gradient sign change rate of adjacent training batches and the preset change rate exceeds 2%, the learning rate warm restart cycle of the nnU-Net model is reduced by 2 epochs for each 1% exceeding, for example, when the difference between the gradient sign change rate of adjacent training batches and the preset change rate is 5%, the learning rate warm restart cycle of the current nnU-Net model is 60 epochs, and the learning rate warm restart cycle of the reduced nnU-Net model is 60x0.9-2x3=48 epochs.

[0114] In implementation, during a long training process, model parameters may exhibit some degradation phenomena, and some parameters may become too large or too small, causing gradient disappearance or explosion, thereby affecting the stability of the loss function. By reducing the learning rate warm restart cycle of the nnU-Net model, the learning rate can be timely restored to a larger value, allowing the model to more actively adjust parameters and adapt to changes in data distribution, which helps to overcome parameter degradation problems and further improves the accuracy of cervical cancer MRI image segmentation.

[0115] Embodiment 1

[0116] The patient information in the training set and the validation set of the nnU-Net model is shown in the following table:

[0117]

[0118]

[0119] Among them, a total of 151 patients were included in the study, and there was no significant difference (p<0.05) in age distribution, pathological type, FIGO stage, maximum tumor diameter, smoking history and HPV infection status, age of first sexual intercourse, number of pregnancies, number of deliveries, number of abortions, etc. between the training set and the test set.

[0120] The nnU-Net automatic delineation model constructed in this study was evaluated based on different evaluation indicators with manual contouring as the standard, and the results are shown in the following table:

[0121] nn-U-Net-2D nn-U-Net-3D DSC 0.7484 0.8181 IoU 0.6367 0.7096 Accuracy 99.53％ 99.73％ Precision 92.18％ 92.09％ Recall 64.91％ 82.52％

[0122] The evaluation indexes include a Dice coefficient, an intersection over union (IoU), a pixel accuracy (Accuracy), a precision, and a recall.

[0123] Specifically, the formula of the Dice coefficient is:

[0124]

[0125] Specifically, the formula of the intersection over union (IoU) is:

[0126]

[0127] Specifically, the formula of the pixel accuracy (Accuracy) is:

[0128]

[0129] Specifically, the formula of the precision is:

[0130]

[0131] In summary, the above two experimental results sufficiently and powerfully prove the effectiveness of the nnU-Net-2D and nnU-Net-3D models in the segmentation of cervical cancer tumors in magnetic resonance images. Through careful analysis of the presented data, it can be clearly found that nnU-Net-3D is overall superior to the 2D model in the segmentation of cervical cancer MRI lesions, especially in improving the detection sensitivity (Recal l) and spatial consistency (DSC / IoU) of tumor infiltration areas. This result verifies the importance of three-dimensional contextual information for accurate segmentation of cervical cancer and provides key technical support for the development of automated tools for clinical staging, target delineation, and efficacy evaluation.

[0132] Thus far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the accompanying drawings, but it is readily understood by those skilled in the art that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without deviating from the principles of the present application, and the technical solutions after these changes or replacements will all fall within the protection scope of the present application.

Claims

1. A cervical cancer MRI image tumor segmentation method based on an nnU-Net network, characterized by, The method comprises the following steps: Collecting a cervical cancer MRI image of a cervical cancer patient, removing invalid background from the cervical cancer MRI image, performing resampling processing and edge slice to output an optimized image; Extracting image features from the optimized image, training an initial model according to the image features to output an nnU-Net model, updating the nnU-Net model by calculating the gradient of each parameter compared to the loss function, and using the nnU-Net model to automatically segment the cervical cancer MRI image to output a tumor region image and a non-tumor region image; Determining whether the accuracy of the cervical cancer MRI image segmentation meets the requirements based on the signal-to-noise ratio of the optimized image; If the accuracy of the cervical cancer MRI image segmentation does not meet the requirements, adjusting the edge slice overlap rate, or determining whether the effectiveness of the nnU-Net model training meets the requirements based on the recall rate of the nnU-Net model; If the effectiveness of the nnU-Net model training does not meet the requirements, adjusting the batch training size of the preprocessed data, or adjusting the learning rate warm restart period of the nnU-Net model based on the gradient sign change rate of adjacent training batches; The increase rate of the edge slice overlap rate is determined by the difference between the signal-to-noise ratio of the optimized image and a preset first signal-to-noise ratio; When the difference between the signal-to-noise ratio of the optimized image and the preset first signal-to-noise ratio is within 2dB, the edge slice overlap rate is increased to 1.2 times the original value, and when the difference exceeds 2dB, the edge slice overlap rate is increased by 2% for every 1dB increase based on the 1.2 times increase; The decrease rate of the batch training size of the preprocessed data is determined by the difference between the recall rate of the nnU-Net model and a preset first recall rate; When the difference between the recall rate of the nnU-Net model and the preset first recall rate is within 0.1, the batch training size of the preprocessed data is reduced to 0.9 times the original value, and when the difference exceeds 0.1, the batch training size of the preprocessed data is reduced by 3 for every 0.05 increase based on the 0.9 times reduction; The decrease rate of the learning rate warm restart period of the nnU-Net model is determined by the difference between the gradient sign change rate of adjacent training batches and a preset change rate; When the difference between the gradient sign change rate of adjacent training batches and the preset change rate is within 2%, the learning rate warm restart period of the nnU-Net model is reduced to 0.9 times the original value, and when the difference exceeds 2%, the learning rate warm restart period of the nnU-Net model is reduced by 2 epochs for every 1% increase based on the 0.9 times reduction.

2. The cervical cancer MRI image tumor segmentation method based on the nnU-Net network according to claim 1, characterized in that, Determining the accuracy of the cervical cancer MRI image segmentation comprises: Comparing the signal-to-noise ratio of the optimized image with a preset second signal-to-noise ratio; If the signal-to-noise ratio of the optimized image is less than or equal to the second signal-to-noise ratio, it is determined that the accuracy of the cervical cancer MRI image segmentation does not meet the requirements.

3. The cervical cancer MRI image tumor segmentation method based on the nnU-Net network according to claim 2, characterized in that, determining the effectiveness of the nnU-Net model training, comprising: comparing the signal-to-noise ratio of the optimized image with the preset first signal-to-noise ratio and the preset second signal-to-noise ratio, respectively; if the signal-to-noise ratio of the optimized image is greater than the preset first signal-to-noise ratio and less than or equal to the preset second signal-to-noise ratio, it is preliminarily determined that the effectiveness of the nnU-Net model training does not meet the requirements, and whether the effectiveness of the nnU-Net model training meets the requirements is determined according to the recall rate of the nnU-Net model.

4. The cervical cancer MRI image tumor segmentation method based on the nnU-Net network according to claim 3, characterized in that, adjusting the edge slice overlap rate, comprising: comparing the signal-to-noise ratio of the optimized image with the preset first signal-to-noise ratio; if the signal-to-noise ratio of the optimized image is less than the preset first signal-to-noise ratio, the edge slice overlap rate is increased.

5. The cervical cancer MRI image tumor segmentation method based on the nnU-Net network according to claim 4, characterized in that, adjusting the batch training size of the preprocessed data, comprising: comparing the recall rate of the nnU-Net model with the preset first recall rate and the preset second recall rate, respectively; if the recall rate of the nnU-Net model is greater than the preset second recall rate, it is determined that the effectiveness of the nnU-Net model training meets the requirements; if the recall rate of the nnU-Net model is less than or equal to the preset first recall rate, it is preliminarily determined that the stability of the loss function does not meet the requirements, and the stability of the loss function is determined according to the gradient sign change rate of adjacent training batches; if the recall rate of the nnU-Net model is greater than the preset first recall rate and less than or equal to the preset second recall rate, the batch training size of the preprocessed data is reduced.

6. The cervical cancer MRI image tumor segmentation method based on the nnU-Net network according to claim 5, characterized in that, adjusting the learning rate warm restart period of the nnU-Net model, comprising: comparing the gradient sign change rate of adjacent training batches with the preset change rate; if the gradient sign change rate of adjacent training batches is greater than the preset change rate, it is determined that the stability of the loss function does not meet the requirements, and the learning rate warm restart period of the nnU-Net model is reduced.

7. The cervical cancer MRI image tumor segmentation method based on the nnU-Net network according to claim 6, characterized in that, the gradient sign change rate of adjacent training batches is the ratio of the number of elements with changed gradient signs in adjacent two batches to the total number of parameter elements.

Images