A craniopharyngioma image segmentation method based on a morphological perception hierarchical information bottleneck network

Abstract

The present application relates to the field of artificial intelligence and medical image processing, and specifically designs a craniopharyngioma image segmentation method based on morphological perception hierarchical information bottleneck network, the method is: acquiring MRI image data of craniopharyngioma patient, pre-processing MRI image data to obtain standardized input image block;Morphological perception hierarchical information bottleneck network MaHIB-Net is constructed;The non-parametric estimation of KL divergence is realized by using kernel density estimation to HIB module, latent feature samples are generated, the probability density of latent feature distribution is calculated, and HIB regularization term is obtained;The total MA loss is obtained by weighted summation;The total loss function of the network is constructed, and the total loss function is the superposition of segmentation loss, weighted sum of HIB regularization term and weighted sum of total MA loss;The segmentation model is trained;The craniopharyngioma MRI image to be segmented is input into the trained segmentation model, and the segmentation result is output, the craniopharyngioma segmentation precision of the present application is high, the field generalization ability is strong, the feature modeling is more flexible and can guarantee the clinical rationality.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence and medical image processing, and particularly relates to a craniopharyngioma image segmentation method based on a morphological perception hierarchical information bottleneck network. BACKGROUND

[0002] Existing craniopharyngioma automatic segmentation technology faces the following core challenges:

[0003] 1. High precision requirement and data scarcity: The incidence of craniopharyngioma is low, resulting in a scarcity of large-scale, expert-annotated MRI datasets; deep learning models rely on a large amount of data for training, and limited data can easily cause model overfitting, making it difficult to adapt to the diverse shapes of tumors such as cystic, solid, and calcified, and unable to guarantee pixel-level segmentation accuracy.

[0004] 2. Insufficient generalization ability in multi-center data domain drift: In clinical practice, patient MRI images are usually obtained from different medical institutions. Due to differences in scanner manufacturers, models, field strengths, acquisition sequence parameters, resolutions, contrast weighting, image post-processing techniques, and patient populations, there is inherent "domain shift" between multi-center data. When a model trained in one center is applied to data from another unseen center, its performance will significantly decrease, severely limiting its clinical applicability and reliability.

[0005] 3. Technical limitations of existing models: Although traditional information bottleneck (IB) methods provide a theoretical framework for learning representations with maximum information and minimum complexity, their variational implementation (VIB) in deep learning is usually based on the assumption that latent variables follow a simple parameter distribution (such as Gaussian distribution). This fixed and restrictive parameter assumption may not accurately capture the complex, non-Gaussian true feature distribution in medical image data, affecting the flexibility and expressiveness of model learning, resulting in suboptimal representation learning results.

[0006] Existing purely data-driven segmentation models may not generate clinically reasonable or anatomically consistent segmentation results, even if they achieve mathematical optimality on the training set;

[0007] Existing advanced models such as nnU-Net, TransUNet, and SwinUNETR perform well in general medical image segmentation tasks, but still have limitations in craniopharyngioma segmentation technology.

[0008] In view of the above background, there is an urgent need to develop an automatic craniopharyngioma segmentation method that can simultaneously solve the problems of high precision, data scarcity, domain drift, and clinical reasonableness. SUMMARY

[0009] In order to solve the problems of insufficient precision, sensitivity to data heterogeneity, poor generalization ability, inflexible modeling of potential feature distribution and lack of clinical rationality existing in the existing craniopharyngioma segmentation, the present application provides a craniopharyngioma image segmentation method based on a morphologically perceived hierarchical information bottleneck network, which has high craniopharyngioma segmentation precision, overcomes the problems of model overfitting and insufficient segmentation precision caused by data scarcity, has strong domain generalization ability, is more flexible in feature modeling and can guarantee clinical rationality.

[0010] The technical scheme of the present application is as follows:

[0011] A craniopharyngioma image segmentation method based on a morphologically perceived hierarchical information bottleneck network, comprising the following steps:

[0012] Step 1: Obtain MRI image data of a craniopharyngioma patient, pre-process the MRI image data to obtain standardized input image blocks;

[0013] Step 2: Construct a morphologically perceived hierarchical information bottleneck network MaHIB-Net, which takes U-Net as the backbone and includes an encoder, a decoder, a hierarchical information bottleneck HIB module and a morphologically perceived MA loss component;

[0014] The HIB module is embedded before the downsampling operation of the encoder and is used to progressively refine and compress the multi-scale features extracted by the encoder; the MA loss component acts on the prediction probability map output by the decoder and is used to constrain the clinical rationality of the segmentation result;

[0015] Step 3: Use kernel density estimation to realize non-parametric estimation of KL divergence for the HIB module, generate latent feature samples through Monte Carlo sampling, calculate the probability density of the latent feature distribution in combination with a Gaussian kernel function and an adaptive bandwidth, and then obtain an HIB regularization term;

[0016] Step 4: Obtain the total MA loss by weighted summation;

[0017] Step 5: Construct a network total loss function, which is the superposition of the segmentation loss, the weighted sum of the HIB regularization term and the weighted sum of the total MA loss;

[0018] Step 6: Train the MaHIB-Net using the pre-processed MRI image data, minimize the total loss function through an optimizer, and obtain a trained segmentation model;

[0019] Step 7: Input the craniopharyngioma MRI image to be segmented into the trained segmentation model, and output the segmentation result of the craniopharyngioma and the surrounding key anatomical structures.

[0020] Preferably, the preprocessing includes image normalization, resampling to a uniform resolution, and random cropping to 128x128x128 voxel image patches.

[0021] Preferably, the number of HIB modules is 4, which are embedded before the deepest four down-sampling operations of the U-Net encoder respectively; each HIB module includes two convolutional layers, followed by an Instance Normalization layer and a LeakyReLU activation function, the input is the feature map of the k-th layer of the encoder, and the output is the compressed latent representation.

[0022] Preferably, the number of samples S of the Monte Carlo sampling is 5;

[0023] The bandwidth of the Gaussian kernel function is adaptively determined according to the Silverman rule, which calculates the bandwidth value based on the dimension and variance of the latent sample; the estimation formula of the KL divergence is:

[0024] ;

[0025] wherein is a standard multivariate Gaussian prior distribution , is the basic sampling distribution generated by the internal encoder of the HIB module, is the probability density estimated by the KDE, and S represents the number of samples of the Monte Carlo sampling.

[0026] Preferably, the MA loss component includes a boundary smoothing loss, a component continuity loss, and an anatomy exclusion loss.

[0027] wherein the formula of the boundary smoothing loss is:

[0028] ;

[0029] wherein is a category set that needs to be forced to smooth the boundary, the category set includes a tumor category, is a voxel set, is the total number of voxels, is the category predicted probability at voxel , is a spatial gradient operator approximated by finite differences, and the weight coefficient of the boundary smoothing loss is set to 0.1 in the experiment.

[0030] The component continuity loss The formula of the segmentation loss is:

[0031] ;

[0032] wherein, is the intensity of the original MRI image at voxel , is the average intensity of voxels predicted as tumor class; is a small constant for numerical stability, and the weight coefficient of the component continuity loss is set to 0.1 in the experiment ; is the predicted probability of tumor class at voxel .

[0033] The formula of the anatomical exclusion loss is:

[0034] ;

[0035] wherein, is the set of critical structure classes that should avoid overlapping with tumor, and the weight coefficient of the anatomical exclusion loss is set to 0.1 is the predicted probability of tumor class at voxel .

[0036] Preferably, in step 5, the segmentation loss is the combination of Dice loss and cross-entropy loss, and the formula of the total loss function is:

[0037] ;

[0038] wherein, is the number of embedded HIB modules, is the compression strength hyperparameter of each HIB module, and is the global weight balancing the relative contribution of HIB regularization and MA loss, is the segmentation loss;

[0039] wherein, the total MA loss is ;

[0040] wherein, is the boundary smoothing loss, is the weight coefficient of the boundary smoothing loss, and the component continuity loss , is the weight coefficient of the component continuity loss, is the anatomical exclusion loss, is the weight coefficient of the anatomical exclusion loss. ​

[0041] Preferably, the optimizer is an Adam optimizer, the initial learning rate is 3x10 -4 , the poly learning rate decay strategy is adopted, the training round is 1000 epochs, the weight decay is 3x10 -5 , and the batch size is 2; during the training process, data enhancement techniques such as random rotation, scaling, elastic deformation and gamma enhancement are adopted.

[0042] Preferably, the key anatomical structure around the craniopharyngioma specifically refers to the pituitary, sphenoid sinus, brain parenchyma, suprasellar cistern, ventricle and background, and a total of 7 anatomical categories are segmented including the tumor.

[0043] Compared with the prior art, the present application has the following beneficial effects:

[0044] (1) Significantly improve the segmentation accuracy and key region delineation ability: MaHIB-Net achieves the highest Dice similarity coefficient (DSC) and intersection over union (IoU) score in the craniopharyngioma segmentation task, and is superior to all the compared SOTA methods. Especially in the complex anatomical regions such as pituitary, a substantial leap in accuracy is achieved. The model generates smoother, more anatomically reasonable boundaries, and significantly reduces false positives and false negatives.

[0045] (2) Excellent field generalization ability and robustness: The verification on the external dataset composed of three unseen medical centers shows that MaHIB-Net has excellent generalization ability and robustness to field drift. It maintains significantly higher performance on cross-center data, and the performance degradation is much lower than the baseline method. This greatly enhances the potential of reliable deployment of the model in different clinical environments.

[0046] (3) Flexible and expressive latent feature modeling: The introduction of KDE for information bottleneck regularization overcomes the limitation of traditional VIB on Gaussian distribution, allowing the model to more flexibly model complex, non-Gaussian latent feature distribution. This enables the model to learn more expressive latent representations, more accurately capturing subtle and critical features in medical image data, and improves the flexibility and effectiveness of feature learning.

[0047] (4) Effectively alleviate the risk of data scarcity and overfitting: The hierarchical feature extraction and compression mechanism of the HIB module, combined with the introduction of domain knowledge by the MA constraint, jointly act on the limited data scenario. This not only reduces the risk of model overfitting, but also improves the robustness of the model in complex craniopharyngioma morphology, thereby improving the reliability of the model in practical applications.

[0048] (5) Ensure the clinical rationality and anatomical consistency of the segmentation results: The MA loss component directly embeds expert anatomical and morphological prior knowledge into the training process, effectively guiding the model to generate segmentation results that are accurate at the voxel level, clinically reasonable and consistent with biological laws in global morphology and topology. This is crucial for downstream clinical applications such as surgical navigation and radiotherapy planning, avoiding clinically unreasonable segmentation that may be produced by purely data-driven models.

[0049] (6) Faster model convergence speed and training stability: The training curve shows that MaHIB-Net exhibits a steeper loss curve descent trend and a consistently higher validation DSC score, indicating faster convergence speed and better training stability of the model.

[0050] (7) The invention provides an innovative technical solution for precise segmentation of craniopharyngiomas, which can be directly applied to neurosurgical clinical work and promote the landing transformation of medical image artificial intelligence. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a flowchart of the method of the invention;

[0052] Figure 2 is an MRI example of craniopharyngioma and key anatomical structures;

[0053] Figure 3 is a model performance comparison curve during training;

[0054] Figure 4 is a conceptual MaHIB-Net architecture diagram;

[0055] Figure 5 is a visual comparison of segmentation results and golden standard GD on external data sets. DETAILED DESCRIPTION

[0056] The invention will be described in detail below in conjunction with the drawings and specific embodiments.

[0057] Referring to Figure 1 and 4 , a craniopharyngioma image segmentation method based on a morphologically perceived hierarchical information bottleneck network includes the following steps:

[0058] Step 1: Obtain MRI image data of a craniopharyngioma patient, pre-process the MRI image data to obtain standardized input image blocks;

[0059] Step 2: Construct a morphological perception hierarchical information bottleneck network MaHIB-Net, which takes U-Net as the backbone, including an encoder, a decoder, a hierarchical information bottleneck HIB module and a morphological perception MA loss component;

[0060] Wherein, the HIB module is embedded before the downsampling operation of the encoder, for progressive refinement and compression of the multi-scale features extracted by the encoder; the MA loss component acts on the prediction probability map output by the decoder, for constraining the clinical rationality of the segmentation result;

[0061] Step 3: The HIB module adopts kernel density estimation to realize non-parametric estimation of KL divergence, generates latent feature samples through Monte Carlo sampling, calculates the probability density of latent feature distribution combined with Gaussian kernel function and adaptive bandwidth, and then obtains the HIB regularization term;

[0062] Step 4: Obtain the total MA loss by weighted summation;

[0063] Step 5: Construct the total loss function of the network, which is the superposition of the segmentation loss, the weighted sum of the HIB regularization term and the total MA loss;

[0064] Step 6: Train MaHIB-Net using preprocessed MRI image data, minimize the total loss function through the optimizer, and obtain the trained segmentation model;

[0065] Step 7: Input the craniopharyngioma MRI image to be segmented into the trained segmentation model, and output the segmentation result of the craniopharyngioma and the surrounding key anatomical structure.

[0066] In the present application, the classical U-Net is used as the basic backbone network, and the encoder part is composed of multiple consecutive convolution blocks (for example, containing two 3x3 convolution layers, followed by Instance Normalization and LeakyReLU activation function) and downsampling layers (for example, 2x2 max pooling), for step-by-step extraction of multi-scale features. The decoder part contains upsampling layers and convolution blocks, for restoring spatial resolution and generating the final segmentation map.

[0067] In an embodiment of the present application, the number of HIB modules is 4, which are respectively embedded before the deepest four downsampling operations of the U-Net encoder; each HIB module includes two convolution layers, which are connected with Instance Normalization layer and LeakyReLU activation function in turn, the input is the k-th layer feature map of the encoder, and the output is the compressed latent representation.

[0068] Integration position and internal structure of HIB module: the present application integrates the HIB module into the U-Net encoder The HIB module is integrated into multiple stages of the U-Net encoder, preferably at deeper levels in the encoder, e.g., before each down-sampling (pooling) operation. Experimental results show that placing the HIB module before the deepest four stages of the U-Net encoder yields the best average foreground Dice score (from 0.796 to 0.811), which indicates that information compression at a more abstract semantic feature level is most effective.

[0069] Each The module itself is a small neural network with an internal structure containing two convolutional layers followed by Instance Normalization and LeakyReLU activation functions. The input to the module is the feature map at the th layer of the U-Net backbone network , and the output is the compressed latent representation .

[0070] KDE-based KL divergence estimation In each module, the internal encoder generates parameters (e.g., mean and variance ) of a base sampling distribution from the input feature map . This base distribution can still be set as a Gaussian distribution for ease of sampling.

[0071] To estimate the KL divergence, a number of Monte Carlo samples are drawn from this base distribution through the reparameterization trick. In the experiments , a good balance between estimation accuracy and computational cost is provided.

[0072] Next, using these samples and a Gaussian kernel function , as well as a specified bandwidth (which can be adaptively determined based on the Silverman rule, based on the dimension and variance of the latent samples), a non-parametric KDE estimation of the probability density function of is performed, denoted as .

[0073] The KDE-approximated KL divergence is estimated by Monte Carlo integration as a HIB regularization term:

[0074] ;

[0075] where is a standard multivariate Gaussian prior distribution .

[0076] The KDE-based KL estimator is differentiable with respect to parameters , ensuring the whole network can be trained end-to-end. By introducing KDE, the latent posterior distribution learned by the model can capture more complex and detailed distribution shapes than simple Gaussian assumption, thus learning a more expressive feature representation.

[0077] In an embodiment of the present application, the preprocessing includes image normalization, resampling to a uniform resolution, and randomly cropping to an image block of 128x128x128 voxels.

[0078] In an embodiment of the present application, the number of samples S = 5 for Monte Carlo sampling;

[0079] The bandwidth of the Gaussian kernel function is adaptively determined according to the Silverman rule, which calculates the bandwidth value based on the dimension and variance of the latent sample; and the estimation formula of the KL divergence is:

[0080] ;

[0081] where is a standard multivariate Gaussian prior distribution , is a basic sampling distribution generated by an internal encoder of the HIB module, is a probability density estimated by KDE, denotes the number of samples S for Monte Carlo sampling.

[0082] In an embodiment of the present application, the MA loss component includes a boundary smoothing loss, a component continuity loss, and an anatomical exclusion loss.

[0083] wherein the formula of the boundary smoothing loss is:

[0084] ;

[0085] wherein, is a category set requiring forced boundary smoothing, the category set including a tumor category, is a voxel set, is a total number of voxels, is a category predicted probability at a voxel . is the spatial gradient operator by finite difference approximation, the weight coefficient of boundary smoothing loss is set in the experiment ;

[0086] The formula of the component continuity loss is:

[0087] ;

[0088] where, is the intensity of the original MRI image at voxel , is the average intensity (weighted by its predicted probability) of voxels predicted to be of tumor class; is a small constant for numerical stability, the weight coefficient of component continuity loss is set in the experiment ; is the predicted probability of tumor class at voxel .

[0089] The formula of the anatomical exclusion loss is:

[0090] ;

[0091] where, is the set of critical structure classes that should not overlap with the tumor, the weight coefficient of anatomical exclusion loss is , is the predicted probability of tumor class at voxel .

[0092] The MA loss component consists of three parts to guide the model to generate clinically reasonable segmentation outputs. Let be the predicted probability map generated by the network, where is the image dimension, is the number of segmentation classes (including background).

[0093] Boundary smoothing loss ( ): Encourages the segmentation boundary to be smooth, reducing the "pixelated" or "noisy" boundary of the prediction result, which is crucial to ensure the accuracy of key clinical measurements such as volume estimation. It is achieved by penalizing the excessive spatial gradient of the predicted probability map;

[0094] Component continuity / homogeneity loss ( ) aims to promote intensity consistency and spatial continuity within the tumor region, preventing the tumor region from being over fragmented or appearing as scattered "blobs". It encourages such consistency by penalizing the variance of the predicted input image intensities within the tumor region, assuming that tumor components (e.g. cystic or solid portions) should have a relatively consistent intensity distribution in a given MRI sequence;

[0095] anatomy exclusion loss ( ) penalizes unreasonable overlap between the predicted tumor region and key non-tumor anatomy structures (e.g. brain parenchyma, ventricles, and in our protocol also the optic chiasm as part of the suprasellar cistern). This is essentially a soft Dice score to minimize the intersection between the tumor class and the exclusion class.

[0096] In an embodiment of the present application, in step 5, the segmentation loss is a combination of Dice loss and cross-entropy loss; the formula of the total loss function is:

[0097] ;

[0098] wherein, is the number of embedded HIB modules, is the compression strength hyperparameter for each HIB module, and is the global weight balancing the relative contribution of HIB regularization and MA loss, is the segmentation loss;

[0099] wherein, the total MA loss is ;

[0100] wherein, is the boundary smoothing loss, is the weight coefficient of the boundary smoothing loss, the component continuity loss , is the weight coefficient of the component continuity loss, is the anatomy exclusion loss, is the weight coefficient of the anatomy exclusion loss.

[0101] wherein, is the number of embedded HIB modules (the present application preferably ), is the compression strength hyperparameter for each HIB module (relatively stable within the range of to , with an optimal value of about ), and is the global weight balancing the relative contribution of HIB regularization and MA loss (optimal value is around ). The setting of these hyperparameters is determined by preliminary experiments on the validation set.

[0102] In an embodiment of the present application, the optimizer is Adam optimizer, the initial learning rate is 3x10 -4 , the poly learning rate decay strategy is adopted, the training round is 1000 epochs, the weight decay is 3x10 -5 , and the batch size is 2; during the training process, the data enhancement techniques of random rotation, scaling, elastic deformation and gamma enhancement are adopted.

[0103] Referring to Figure 2 , in an embodiment of the present application, the key anatomical structures around the craniopharyngioma specifically refer to the pituitary, sphenoid sinus, brain parenchyma, suprasellar cistern, ventricle and background, including 7 anatomical categories in total.

[0104] Referring to Figure 3 , in the present application, in order to comprehensively evaluate the effectiveness of the present application, the following current state-of-the-art (SOTA) methods are compared in the experiment:

[0105] • nnU-Net (full-resolution configuration): as the SOTA method in the field of medical image segmentation, it realizes high performance through automatic preprocessing, optimized 3D U-Net architecture and model integration. In the experiment, it is used as the baseline because it performs well in various tasks.

[0106] • TransUNet: a hybrid CNN-Transformer architecture that combines the local feature extraction capability of U-Net and the global context modeling of Vision Transformer, aiming to handle long-range dependencies.

[0107] • SwinUNETR: a U-Net variant based on hierarchical Swin Transformer, which uses shift window attention mechanism for efficient long-range dependency modeling, and has been widely used in medical image processing.

[0108] These comparison methods are carried out under the same experimental settings (dataset, training strategy, evaluation index), ensuring the fairness and persuasiveness of the comparison results.

[0109] The segmentation results are shown in Figure 5 , and the visual comparison of the segmentation results on the external dataset with the gold standard GD shows the accuracy of the present application.

[0110] The above merely illustrates the embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, which are made by using the content of the present application specification and drawings, are also included in the patent protection scope of the present application.

Claims

1. A craniopharyngioma image segmentation method based on a morphological perception hierarchical information bottleneck network, characterized in that, The method comprises the following steps: Step 1: obtaining MRI image data of a craniopharyngioma patient, pre-processing the MRI image data to obtain a standardized input image block; Step 2: constructing a morphological perception hierarchical information bottleneck network MaHIB-Net, the network taking U-Net as a backbone, comprising an encoder, a decoder, a hierarchical information bottleneck HIB module and a morphological perception MA loss component; The HIB module is embedded before the downsampling operation of the encoder, and is used for progressive refinement and compression of multi-scale features extracted by the encoder; the MA loss component acts on the prediction probability map output by the decoder, and is used for constraining the clinical rationality of the segmentation result; Step 3: using kernel density estimation to realize non-parametric estimation of KL divergence for the HIB module, generating latent feature samples through Monte Carlo sampling, combining a Gaussian kernel function and an adaptive bandwidth to calculate the probability density of the latent feature distribution, and then obtaining an HIB regularization term; Step 4: obtaining a total MA loss through weighted summation; Step 5: constructing a network total loss function, the total loss function being a superposition of a segmentation loss, a weighted sum of the HIB regularization term and a weighted sum of the total MA loss; Step 6: training the MaHIB-Net using the pre-processed MRI image data, minimizing the total loss function through an optimizer to obtain a trained segmentation model; Step 7: inputting a craniopharyngioma MRI image to be segmented into the trained segmentation model, and outputting a segmentation result of the craniopharyngioma and surrounding key anatomical structures; The sample number S of the Monte Carlo sampling is 5; The bandwidth of the Gaussian kernel function is adaptively determined according to a Silverman rule, the Silverman rule calculating a bandwidth value based on dimensions and variance of potential samples; and the estimation formula of the KL divergence is: ； where is a standard multivariate Gaussian prior distribution , is the base sampling distribution generated by the internal encoder of the HIB module, is the probability density estimated by the KDE, denotes the number of samples of the S Monte Carlo samples.

2. The method of claim 1, wherein the method is a morphological perception hierarchical information bottleneck network-based craniopharyngioma image segmentation method. The pre-processing comprises image normalization, resampling to a unified resolution and randomly cropping an image block of 128x128x128 voxels.

3. The method of claim 1, wherein the method is a morphological perception hierarchical information bottleneck network-based craniopharyngioma image segmentation method. The number of the HIB modules is 4, which are embedded before the deepest four downsampling operations of the U-Net encoder; each HIB module comprises two convolutional layers, followed by an Instance Normalization layer and a LeakyReLU activation function, the input being a feature map of the kth layer of the encoder, and the output being a compressed latent representation.

4. The method of claim 1, wherein the method is a morphological perception hierarchical information bottleneck network-based craniopharyngioma image segmentation method. The MA loss component comprises a boundary smoothing loss, a component continuity loss and an anatomical exclusion loss; In the formula, the boundary smoothness loss is The formula is: ； wherein, is a set of classes requiring forced boundary smoothing, the set of classes comprising a tumor class, is a set of voxels, is a total number of voxels, is a class a predicted probability at a voxel , is a spatial gradient operator approximated by finite differences; Loss of component continuity The formula is: ； wherein, is the intensity of the voxel at the original MRI image, is the average intensity of the voxels predicted to be of the tumor class; is a small constant for numerical stability; is the tumor class the predicted probability at the voxel . The anatomical exclusion loss The formula is: ； wherein, is a set of key structure classes that should be avoided from overlapping with the tumor, the anatomical exclusion loss weight coefficient , is a tumor class the predicted probability at a voxel .

5. The method of claim 1, wherein the method is a morphological perception hierarchical information bottleneck network-based rathke cleft cyst image segmentation method. In step 5, the segmentation loss is a combination of Dice loss and cross-entropy loss; the total loss function The formula is: ； wherein, is the number of embedded HIB modules, is the compression strength hyperparameter for each HIB module, and is the global weight balancing the relative contribution of HIB regularization and MA loss, is the segmentation loss; wherein the total MA loss is ; wherein, is a boundary smoothness loss, is a weight coefficient of the boundary smoothness loss, component continuity loss , is a weight coefficient of the component continuity loss, is an anatomical exclusion loss, is a weight coefficient of the anatomical exclusion loss.

6. The method of claim 1, wherein the method is a morphological perception hierarchical information bottleneck network-based craniopharyngioma image segmentation method. The optimizer is an Adam optimizer, the initial learning rate is 3x10 -4 , a poly learning rate decay strategy is adopted, the training round is 1000 epochs, the weight decay is 3x10 -5 , and the batch size is 2; during the training process, data enhancement techniques such as random rotation, scaling, elastic deformation and gamma enhancement are adopted.

7. The method of claim 1, wherein the method is a morphological perception hierarchical information bottleneck network-based craniopharyngioma image segmentation method. The surrounding key anatomical structures of the craniopharyngioma specifically refer to the pituitary, the sphenoid sinus, the brain parenchyma, the suprasellar cistern, the ventricle and the background, and a total of 7 anatomical categories are segmented including the tumor.

Images