Trigeminal neuralgia segmentation and classification system based on Merkel cavity

Abstract

The invention provides a trigeminal neuralgia segmentation and classification system based on a Merkel cavity, and relates to the technical field of disease classification. Comprising a trigeminal nerve magnetic resonance imaging segmentation subsystem which is used for automatically segmenting a medical image which is obtained based on magnetic resonance imaging and contains a trigeminal nerve Merkel cavity so as to obtain a segmentation result of the Merkel cavity, and determining a region of interest for classification and use according to the segmentation result; and the trigeminal neuralgia identification and classification subsystem is used for performing trigeminal neuralgia and normal identification and classification on the tested object based on the quantitative characteristics of the segmentation result, and outputting a classification result and probability information. The problems that in the prior art, in trigeminal neuralgia iconography application, manual segmentation is mainly adopted, and segmentation precision and small sample generalization are insufficient are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of disease classification, and in particular to a Meckel's cave-based trigeminal neuralgia segmentation and classification system. BACKGROUND

[0002] Trigeminal neuralgia (TN) is a chronic neurological disorder characterized by sudden, electric-shock-like pain in the areas of the face innervated by the trigeminal nerve. The pain is often described by patients as "unbearable", severely affecting daily life and work. Current clinical diagnosis is mainly based on patient symptom description and physician experience, which is a subjective diagnosis method that can easily lead to misdiagnosis or missed diagnosis. Although antiepileptic drugs such as carbamazepine are the first choice for treatment, long-term use often leads to drug resistance and significant side effects, which makes more than half of the patients ultimately seek surgical treatment.

[0003] Among existing surgical treatment methods, microvascular decompression (MVD) is considered the gold standard for treating trigeminal neuralgia. However, this surgery requires craniotomy, which is not only traumatic but also carries the risk of serious complications such as cerebrospinal fluid leakage and intracranial infection. More importantly, for patients without clear evidence of vascular compression, the effectiveness of this surgery is often less than satisfactory. In contrast, percutaneous balloon compression (PBC) has become an important alternative treatment option due to its minimally invasive and simple operation. However, PBC also has obvious limitations, with a postoperative recurrence rate as high as 20-30%, and often leading to serious complications such as facial numbness.

[0004] Recent studies have found that the effectiveness of PBC is closely related to the morphological characteristics of Meckel's Cave (MC). Clinical data shows that patients with larger MC volume (> 484.3 mm³) have significantly lower postoperative pain relief rates, only 51.8%, while patients with smaller MC volume can achieve a relief rate of 86.4%. In addition, the morphological characteristics of MC are also important influencing factors, patients with spherical MC with a flatness exceeding 0.402 have a significantly increased risk of recurrence, which may be because the pressure distribution within the spherical MC is uneven, resulting in insufficient compression of the trigeminal ganglion.

[0005] Traditional imaging analysis mainly focuses on the contact relationship between the trigeminal nerve and blood vessels in the cisternal segment, while ignoring the importance of MC morphological features. In fact, features such as the volume, surface area, and texture heterogeneity of MC play a key role in the pathogenesis and treatment response of TN. For example, the flattening of MC changes the mechanical compression distribution of the trigeminal ganglion, thereby directly affecting the treatment effect of PBC. Although imaging omics technology can quantify the morphological heterogeneity of MC by extracting high-order features (such as gray level co-occurrence matrix GLCM, wavelet transform, etc.), current research mostly relies on manual segmentation, which is not only inefficient but also prone to subjective bias.

[0006] The main challenges of the current technology include: first, in terms of segmentation accuracy, traditional models such as U-Net have limited generalization ability for small sample medical images (such as MC), especially in difficult-to-segment regions such as the edge of the nerve branch, and the Dice coefficient is often less than 0.75; second, existing research often only focuses on a single function (segmentation or classification), lacks a closed-loop system that combines segmentation and classification, and greatly limits clinical practicability; finally, the image acquisition parameters and segmentation protocols used by different medical institutions are not unified, which seriously hinders the cross-center promotion and application of the model. SUMMARY

[0007] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a Meckel's cave-based trigeminal neuralgia segmentation and classification system, which solves the problems of manual segmentation as the main method, insufficient segmentation accuracy and small sample generalization in the current technology in the application of trigeminal neuralgia imaging.

[0008] To achieve the above-mentioned purpose, the present application provides the following scheme:

[0009] A Meckel's cave-based trigeminal neuralgia segmentation and classification system, comprising:

[0010] A trigeminal nerve magnetic resonance imaging segmentation subsystem for automatically segmenting medical images containing Meckel's cave of the trigeminal nerve based on magnetic resonance imaging to obtain segmentation results of Meckel's cave, and determining a region of interest for classification based on the segmentation results;

[0011] A trigeminal neuralgia differential classification subsystem for differentiating and classifying subjects with trigeminal neuralgia and normal based on quantitative features of the segmentation results, and outputting classification results and probability information.

[0012] Preferably, the training process of the trigeminal nerve magnetic resonance imaging segmentation subsystem comprises:

[0013] S1: Collect and label image data containing Meckel's cave of the left and right trigeminal nerves;

[0014] S2: preprocessing the data and dividing the training set, validation set and test set;

[0015] S3: selecting a deep learning model and parameters for Meckel's cavity segmentation;

[0016] S4: based on the morphological structure characteristics of Meckel's cavity and surrounding tissues, the region is separated, and a combined loss function containing Dice loss and cross-entropy loss is used to optimize the segmented region;

[0017] S5: randomly sampling image blocks from the training set, performing data augmentation and normalization processing, and inputting the model;

[0018] S6: generate segmentation results by forward propagation, calculate loss and update model parameters by back propagation;

[0019] S7: repeat steps S5 and S6 until the preset training iteration condition is reached;

[0020] S8: evaluate the segmentation performance indicators based on the validation set and select the optimal model;

[0021] S9: evaluate the segmentation performance of the optimal model on the test set.

[0022] Preferably, the construction process of the trigeminal neuralgia identification and classification subsystem comprises:

[0023] T1: build a data set and label the categories according to clinical diagnosis;

[0024] T2: use the segmentation subsystem to automatically segment Meckel's cavity, and extract morphological and intensity quantitative features of Meckel's cavity based on the segmentation mask;

[0025] T3: standardize the features and divide the training set, validation set and test set;

[0026] T4: select a classification model and optimize the hyperparameters;

[0027] T5: train and iteratively update the classification model;

[0028] T6: evaluate the classification performance on the test set and generate classification results.

[0029] Preferably, the inference of the segmentation subsystem comprises:

[0030] P1: convert the test sample image to NIfTI format and organize it according to the predetermined specification;

[0031] P2: automatically load the preprocessing plan, perform resampling and standardization consistent with the training phase;

[0032] P3: adopt sliding window to perform ensemble inference on the model set obtained by k-fold cross-validation, obtain a probability map and reconstruct by splicing in a Gaussian weighting manner;

[0033] P4: perform argmax on the probability map to generate initial labels, and remove false positive regions by post-processing such as connected component analysis to obtain binary segmentation;

[0034] P5: superimpose the segmentation result on the original image and perform three-dimensional reconstruction.

[0035] Preferably, the quantitative features extracted in step T2 include:

[0036] Morphological features: volume, surface area, maximum diameter, equivalent spherical diameter, and voxel number;

[0037] Intensity features: mean, standard deviation, maximum value, and minimum value of signal intensity in the segmented region;

[0038] Asymmetry features: ratio or difference of the above features of the left and right sides of the Meckel's cavity.

[0039] Preferably, the classification subsystem includes the following steps in the inference stage:

[0040] C1: call the segmentation subsystem to obtain the Meckel's cavity segmentation of the test sample and the corresponding features;

[0041] C2: perform consistent Z-Score transformation on the test features based on the standardized parameters saved in the training stage;

[0042] C3: input the standardized features into the optimal classification model to obtain the prediction probability of each category;

[0043] C4: output the classification probability, the discrimination result, and generate an analysis report including the confusion matrix, ROC curve, and feature importance ranking.

[0044] Preferably, the segmentation performance evaluation indicators include Dice coefficient, recall rate, ROC-AUC, Hausdorff distance, and precision; and the segmentation performance indicators include:

[0045] AUC, sensitivity, specificity, accuracy, precision, and F1 score.

[0046] The present application discloses the following technical effects:

[0047] The application provides a Meckel's cave-based trigeminal neuralgia segmentation and classification system, comprising: a trigeminal nerve magnetic resonance imaging segmentation subsystem for automatically segmenting a medical image containing a Meckel's cave of a trigeminal nerve acquired based on magnetic resonance imaging to obtain a segmentation result of the Meckel's cave and determine a region of interest for classification; and a trigeminal neuralgia differential classification subsystem for differentially classifying a subject into trigeminal neuralgia and normal based on quantitative features of the segmentation result and outputting a classification result and probability information. The application realizes high-precision segmentation of the Meckel's cave through an nnUNet framework and a hybrid loss function, solves the problem of insufficient generalization ability of traditional methods in small sample medical images, realizes automatic classification (AUC is greater than or equal to 0.90) of trigeminal neuralgia based on morphological features of the Meckel's cave and a machine learning algorithm, avoids subjective bias in traditional clinical diagnosis, provides a personalized surgical plan suggestion (such as adjusting the balloon pressure according to the Meckel's cave morphology) for a trigeminal neuralgia patient, assists in efficacy evaluation and improves the prognosis of the patient, and solves the problem of non-uniform image acquisition parameters of different medical institutions through standardized preprocessing and model design, facilitating popularization. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0049] Figure 1 A Meckel's cave-based trigeminal neuralgia segmentation and classification system structure schematic diagram is provided for the embodiments of the present application.

[0050] Figure 2 A normalization flowchart is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0053] As Figure 1As shown, the present application provides a Merkel cavity-based trigeminal neuralgia segmentation and classification system, comprising:

[0054] a trigeminal nerve magnetic resonance imaging segmentation subsystem for automatically segmenting medical images containing Merkel cavities of the trigeminal nerve acquired based on magnetic resonance imaging to obtain segmentation results of the Merkel cavities and determine a region of interest for classification based on the segmentation results;

[0055] a trigeminal neuralgia differential classification subsystem for differentially classifying a subject as having trigeminal neuralgia or being normal based on quantitative features of the segmentation results and outputting classification results and probability information.

[0056] Specifically, the above subsystems correspond to the following workflow: in the training phase, input the homologous training data of the test image (T2 original image and trigeminal Merkel cavity annotation), train the Merkel cavity segmentation algorithm based on the preset training parameters to generate a segmentation model file, and at the same time align the segmentation results with the original image and use them to generate images for 3D rendering to check the effect; in the test phase, import the T2 original image as the test image, call the trained segmentation model file to automatically segment the Merkel cavity, output the segmentation results and map them to the original image to realize visualization; thereafter, in the trigeminal neuralgia classification system, the training phase extracts features from the T2 original image and the corresponding trigeminal Merkel cavity annotation (calibrated by the segmentation results) and trains a classification model file for patients with trigeminal neuralgia and normal people, and a supervision signal is established with reference to the clinical gold standard label; in the test phase, extract classification features from the above segmentation results, load the classification model file to output the probability of patients with trigeminal neuralgia and the probability of normal people, and give the final output result (patient / normal) and optional visualization of the comparison between patients with trigeminal neuralgia and normal people.

[0057] Further, the training process of the trigeminal nerve magnetic resonance imaging segmentation subsystem comprises:

[0058] S1: collect and label image data containing left and right Merkel cavities of the trigeminal nerve;

[0059] S2: pre-process the data and divide it into a training set, a validation set and a test set;

[0060] S3: select a deep learning model and parameters for Merkel cavity segmentation;

[0061] S4: separate the regions based on the morphological structural features of the Merkel cavity and the surrounding tissues, and optimize the segmented regions using a combined loss function containing Dice loss and cross-entropy loss;

[0062] S5: randomly sample image blocks from the training set, perform data augmentation and normalization processing, and input them into the model;

[0063] S6: generate segmentation results by forward propagation, calculate loss and update model parameters by back propagation;

[0064] S7: repeat steps S5 and S6 until the preset training iteration condition is reached;

[0065] S8: evaluate the segmentation performance indicators based on the validation set and select the optimal model;

[0066] S9: evaluate the segmentation performance of the optimal model on the test set.

[0067] Specifically, first (S1), collect and label the image data of the left and right two Meckel's cavities of the trifacial nerve; then (S2) perform data preprocessing and divide the training set, validation set and test set according to the ratio of 7:1:2; then (S3) select a deep learning model and corresponding parameters according to the task requirements; in (S4), separate the Meckel's cavity and the surrounding tissue according to the morphological structure, and use a combined loss (the total loss is the sum of the Dice loss and the cross-entropy loss) to optimize the segmentation, wherein the Dice loss is calculated based on the predicted foreground probability of each pixel and the true label, and a small amount is introduced to ensure numerical stability, and the cross-entropy loss is calculated based on the pixel-level classification error in the range of all pixels; then (S5) randomly sample samples from the training set, read after automatic correction by nnUNet, and randomly crop image blocks of fixed size and resolution in each sample range according to the set sampling strategy, perform data augmentation and normalization; then (S6) generate Meckel's cavity segmentation results by forward propagation, calculate the deviation and gradient between the combined loss and the manual annotation, and update the model parameters by back propagation; then (S7) repeat S5 and S6 until the entire training set is traversed and the maximum iteration number is reached; in (S8), evaluate the model using the validation set, calculate and compare the DICE coefficient, recall rate, ROC-AUC, Hausdorff distance and precision to select the optimal model; finally (S9) test the optimal model on the test set to obtain the performance indicators such as the DICE coefficient, recall rate, ROC-AUC, Hausdorff distance and precision of the test set.

[0068] Specifically, the combined loss function is:

[0069] ;

[0070] The Dice loss function formula is as follows:

[0071] ;

[0072] The cross-entropy loss (Cross-Entropy Loss) function is as follows:

[0073] ;

[0074] pi represents the probability (range [0, 1]) that the i-th pixel predicted by the model is foreground (Meckel's cavity); gi represents the value of the i-th pixel in the true label (0 or 1); epsilon represents a very small number to ensure numerical stability (to prevent the denominator from being 0); N represents the total number of pixels in the image; i indexes each pixel; p represents the probability that the model predicts that the pixel is foreground. g represents the true label of the pixel (1 if foreground, 0 if background).

[0075] Further, the three-pronged magnetic resonance imaging segmentation system trains a segmentation model according to different images.

[0076] The deep learning model includes: nnUNet, and the designed parameters include a dataset path, a dataset ID, a training fold number, and the like, and the remaining parameters including a learning rate, a patch size, a network structure, a block size, and the like are self-provided by the model and do not need to be adjusted.

[0077] As shown in Figure 2 , Z-Score normalization, the process is as follows:

[0078] a. Foreground mask: First, create a mask of the foreground region based on the resampled label or the image itself (if the label is not complete). This can avoid the huge impact of the background region (a large number of 0 values) on the calculation of intensity statistics.

[0079] b. Calculate the statistical quantity of each image: For each training image, separately calculate the mean and standard deviation (std) of the intensity of the foreground region.

[0080] c. Clip outliers: In order to avoid a small number of extreme outliers (such as metal artifacts and noise spikes in MRI) in the image distorting the entire intensity distribution, nnU-Net will clip the image intensity based on the 0.5% and 99.5% quantiles of the foreground pixel intensity. Any value below the 0.5% quantile is set to the 0.5% quantile, and any value above the 99.5% quantile is set to the 99.5% quantile.

[0081] d. Z-Score normalization: Standardize using the mean and std of each image calculated in step b:

[0082] ;

[0083] where I clipped : the clipped image; mu foreground : the mean of the foreground region of the image; sigma foreground : the standard deviation of the foreground region of the image.

[0084] Further, the inference of the segmentation subsystem includes:

[0085] P1: Convert the test sample image to NIfTI format and organize it according to the predetermined specification;

[0086] P2: Automatically load the preprocessing plan, perform resampling and standardization consistent with the training phase;

[0087] P3: Use sliding window to perform integrated inference on the model set obtained by k-fold cross-validation, obtain probability map and reconstruct by Gaussian weighting method;

[0088] P4: Perform argmax on the probability map to generate initial labels, and remove false positive regions through post-processing such as connected component analysis to obtain binary segmentation;

[0089] P5: Superimpose the segmentation result on the original image and perform three-dimensional reconstruction.

[0090] Specifically, data preparation and preprocessing: convert the original magnetic resonance image of the test sample (without pre-preprocessing bias field correction) to NIfTI format, and store it in the specified directory according to the naming specification required by nnU-Net;

[0091] Automatic preprocessing: call the nnU-Net inference engine, which will automatically load the preprocessing plan generated for the Meckel's cavity dataset, and perform the same preprocessing process as the training phase on the input test image, including:

[0092] a. Resampling: automatically resample the image to the target interval calculated by nnU-Net.

[0093] b. Standardization.

[0094] Model inference and prediction: input the preprocessed image into the optimal model set automatically planned and trained by nnU-Net; the model set includes all n models trained in k-fold cross-validation; the inference process uses sliding window method to automatically process images of any size, specifically:

[0095] a. Sliding window segmentation: according to the optimal patch size (PatchSize) automatically calculated by nnU-Net, generate overlapping image blocks in sliding window manner on the image.

[0096] b. Integrated prediction: each image block is input into the n trained models for forward propagation, and the prediction probability maps of all models are averaged to generate an integrated, high-confidence prediction result.

[0097] c.Result reconstruction: map the prediction results of all image patches back to their original positions, and perform a Gaussian-weighted average of the prediction probabilities in the overlapping regions, finally stitch them into a complete probability map with the same size as the preprocessed image.

[0098] Post-processing and result output: apply argmax operation on the complete probability map obtained by stitching to generate the initial segmentation label map; then, nnU-Net automatically applies its built-in post-processing algorithm (such as based on connected component analysis, retaining the largest several foreground regions) to remove tiny false positive noise points, obtaining the final binary segmentation result.

[0099] Result visualization: superimpose the final segmentation result on the original magnetic resonance image and perform three-dimensional reconstruction rendering for user interactive viewing and evaluation.

[0100] Further, the construction process of the trigeminal neuralgia identification and classification subsystem comprises:

[0101] T1: Construct a dataset and label the categories according to clinical diagnosis;

[0102] T2: Use the segmentation subsystem to automatically segment the Meckel's cave, and extract morphological and intensity quantitative features of the Meckel's cave based on the segmentation mask;

[0103] T3: Standardize the features and divide them into training set, validation set and test set;

[0104] T4: Select a classification model and optimize the hyperparameters;

[0105] T5: Train and iteratively update the classification model;

[0106] T6: Evaluate the classification performance on the test set and generate the classification result.

[0107] Specifically, data set construction and label definition: collect magnetic resonance image data sets containing Meckel's cave, and assign category labels to each sample according to clinical diagnosis results (such as whether there is neurovascular compression, whether it is diagnosed as trigeminal neuralgia, etc.);

[0108] Automatic segmentation and feature extraction of Meckel's cave: the nnU-Net segmentation subsystem automatically segments the Meckel's cave of all images in the dataset; based on the binary mask obtained by segmentation, a series of quantitative morphological and intensity features are automatically extracted, including but not limited to:

[0109] Morphological features: volume, surface area, maximum diameter, equivalent spherical diameter, and voxel number of Meckel's cave;

[0110] Intensity features: mean, standard deviation, maximum, and minimum of the signal intensity of the segmented region within the Meckel's cave;

[0111] Asymmetry features: for bilateral Meckel cavities, calculate the ratio or difference of the above features between left and right sides;

[0112] Dataset splitting and feature standardization: split the extracted feature data matrix and its corresponding class labels into training, validation, and test sets in a certain proportion (e.g., 7:1:2); and perform Z-Score standardization on all features to eliminate dimensional effects;

[0113] Classifier selection and hyperparameter optimization: select a machine learning classification model (such as support vector machine SVM, random forest, XGBoost, or shallow neural network), and use the validation set to optimize the model hyperparameters through grid search or Bayesian optimization;

[0114] Model training and iteration: input the features and labels of the training set into the selected classification model for training, and monitor the performance on the validation set to prevent overfitting; repeat the process until the model converges or reaches optimal performance;

[0115] Model evaluation and selection: evaluate the performance of the final model on the test set, calculate AUC, sensitivity, specificity, accuracy, precision, and F1 score indicators to objectively evaluate its disease discrimination ability.

[0116] Further, the classification subsystem in the inference stage includes:

[0117] C1: call the segmentation subsystem to obtain the Meckel cavity segmentation of the test sample and the corresponding features;

[0118] C2: perform consistent Z-Score transformation on the test features based on the standardized parameters saved during the training phase;

[0119] C3: input the standardized features into the optimal classification model to obtain the prediction probability of each class;

[0120] C4: output the classification probability, discrimination result, and generate an analysis report including the confusion matrix, ROC curve, and feature importance ranking.

[0121] Further, the segmentation performance evaluation indicators include Dice coefficient, recall rate, ROC-AUC, Hausdorff distance, and precision; the segmentation performance indicators include:

[0122] AUC, sensitivity, specificity, accuracy, precision, and F1 score.

[0123] Specifically, the integrated learning model is preferably selected from random forest (Random Forest), gradient boosting machine (such as XGBoost), or support vector machine (SVM).

[0124] Automatic segmentation and feature extraction: input the test sample image to be identified into the nnU-Net segmentation subsystem, and automatically obtain the fine segmentation result of the Meckel's cave; then automatically perform the step "automatic segmentation of the Meckel's cave of all images in the data set" to extract the same quantitative feature vector;

[0125] Feature preprocessing: load the feature standardization parameters (mean and standard deviation) saved in the training phase, and perform the same Z-Score standardization transformation on the extracted test sample feature vector;

[0126] Forward prediction: input the standardized feature vector into the trained optimal classification model, and obtain the prediction probability of belonging to each category through forward propagation;

[0127] Result generation and visualization: output the final classification probability and identification result; and generate a high-level analysis report, including but not limited to a confusion matrix, a ROC curve, and a feature importance ranking chart, to provide quantitative basis for clinical decision-making.

[0128] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0129] The principles and implementation modes of the present application are described by applying specific examples in this paper. The above description of the embodiments is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A trigeminal neuralgia segmentation and classification system based on Merkel's cavity, characterized in that, include: The trigeminal nerve magnetic resonance imaging segmentation subsystem is used to automatically segment medical images containing the Merkel cavity of the trigeminal nerve obtained from magnetic resonance imaging, so as to obtain the segmentation results of the Merkel cavity and determine the region of interest for classification. The trigeminal neuralgia differentiation and classification subsystem is used to differentiate between trigeminal neuralgia and normal conditions in subjects based on the quantitative features of the segmentation results, and outputs the classification results and probability information.

2. The trigeminal neuralgia segmentation and classification system based on Merkel's cavity according to claim 1, characterized in that, The training process for the trigeminal nerve magnetic resonance imaging segmentation subsystem includes: S1: Collect and label imaging data containing the left and right Merkel cavities of the trigeminal nerve; S2: Preprocess the data and divide it into training set, validation set and test set; S3: Select the deep learning model and parameters for Merkel cavity segmentation; S4: Region separation is performed based on the morphological structural features of the Merkel cavity and surrounding tissues, and the segmented region is optimized using a combined loss function including Dice loss and cross-entropy loss; S5: Randomly sample image patches from the training set, perform data augmentation and normalization, and then input them into the model; S6: Generate segmentation results through forward propagation, calculate the loss, and update the model parameters through backpropagation; S7: Repeat steps S5 and S6 until the preset training iteration conditions are met; S8: Evaluate segmentation performance metrics based on the validation set and select the optimal model; S9: Evaluate the segmentation performance of the optimal model on the test set.

3. The trigeminal neuralgia segmentation and classification system based on Merkel's cavity according to claim 1, characterized in that, The construction process of the trigeminal neuralgia differential classification subsystem includes: T1: Construct the dataset and label the categories according to clinical diagnoses; T2: The Merkel cavity is automatically segmented using the segmentation subsystem, and the morphological and intensity quantitative features of the Merkel cavity are extracted based on the segmentation mask; T3: Standardize the features and divide them into training, validation and test sets; T4: Select a classification model and optimize its hyperparameters; T5: Train and iteratively update the classification model; T6: Evaluate classification performance on the test set and generate classification results.

4. A trigeminal neuralgia segmentation and classification system based on Merkel's cavity according to claim 1, characterized in that, The reasoning of the segmentation subsystem includes: P1: Convert the test sample image to NIfTI format and organize it according to the predetermined specifications; P2: Automatically load the preprocessing plan and perform resampling and normalization consistent with the training phase; P3: Using a sliding window, ensemble inference is performed on the model set obtained by k-fold cross-validation to obtain a probability map, which is then spliced ​​and reconstructed in a Gaussian weighted manner. P4: argmax is applied to the probability map to generate initial labels, and false positive regions are removed by post-processing such as connected component analysis to obtain binary segmentation; P5: The segmentation results are overlaid on the original image and then 3D reconstruction is performed.

5. A trigeminal neuralgia segmentation and classification system based on Merkel's cavity according to claim 1, characterized in that, The quantitative features extracted in step T2 include: Morphological characteristics: volume, surface area, maximum diameter, equivalent spherical diameter, number of voxels; Intensity characteristics: mean, standard deviation, maximum, and minimum signal intensity within the segmented region; Asymmetric features: the ratio or difference of the above features between the left and right Merkel cavities.

6. A trigeminal neuralgia segmentation and classification system based on Merkel's cavity according to claim 1, characterized in that, The classification subsystem includes the following in the reasoning phase: C1: Call the segmentation subsystem to obtain the Merkle cavity segmentation and corresponding features of the test sample; C2: Apply a consistent Z-Score transformation to the test features based on the standardized parameters saved during the training phase; C3: Input standardized features into the optimal classification model to obtain the predicted probabilities of each category; C4: Outputs classification probabilities and discrimination results, and generates an analysis report including confusion matrix, ROC curve and feature importance ranking.

7. A trigeminal neuralgia segmentation and classification system based on Merkel's cavity according to claim 1, characterized in that, The segmentation performance evaluation metrics include Dice coefficient, recall, ROC-AUC, Hausdorff distance, and precision; the segmentation performance metrics include: AUC, sensitivity, specificity, accuracy, precision, and F1 score.

Images