Neural tissue identification method based on artificial intelligence

Abstract

The invention provides a nervous tissue identification method based on artificial intelligence, and relates to the technical field of medical intelligent detection. According to the nervous tissue identification method, nervous tissue region annotation is carried out on 452 HE-dyed section images of bladder cancer, pancreatic cancer and porta-bile duct cancer by using QuPath software. 70% is used as a training set, 10% is used as a verification set of a training process, and 20% is used as an independent test set to train the nnUnet model. The whole training process is combined with an active learning strategy to enhance the specificity of the model. By using the algorithm and the deep learning technology, the nerve peripheral infiltration in the pathology is accurately and quickly evaluated, the workload of manual identification is greatly reduced, and the diagnosis accuracy is improved. The machine identification of tumor peripheral nerve invasion has obvious practicability, importance, clinical diversity and actual clinical value, and shows that the deep learning technology improves the potential of tumor detection and diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of medical intelligent detection, and in particular to a neural tissue recognition method based on artificial intelligence. Background Art

[0002] Tumors are new growths formed by the proliferation of local tissue cells. According to the cellular characteristics of the new growths and the degree of harmfulness to the body, tumors are divided into two categories: benign tumors and malignant tumors. Perineural infiltration is an important process in the development of tumors, and it is the malignant spread of tumor cells to the peripheral nerves. When the tumor develops to the late stage, the tumor infiltrates the adjacent nerve tissue, exerts pressure on the nerve structure or destroys the nerve structure. This destruction impairs nerve function and is often a key determinant of poor prognosis for cancer patients. Although perineural infiltration can be manifested in various types of solid tumor sites, it usually affects areas such as the pelvis, head and neck, and retroperitoneal tumors.

[0003] The impact of perineural invasion on a patient's overall health and quality of life is profound. Invasion of sensory nerves can cause paresthesias, including pain, numbness, or tingling. Similarly, invasion of motor nerves can impair muscle function, leading to muscle weakness or limited limb movement. In addition, perineural invasion can cause neurological symptoms such as autonomic dysfunction, neurogenic pruritus, digestive and excretory disorders, and psychological disorders. It is worth noting that the severity of peripheral nerve invasion and the damage caused may vary between individuals, depending on factors such as tumor type, location, extent of invasion, and the individual's overall health status, all of which can significantly affect the patient's quality of life.

[0004] A large number of scientific studies have shown that there is a close correlation between perineural invasion and various tumor characteristics, including invasiveness, metastatic tendency, poor quality of life, and poor prognosis. Therefore, it is crucial to accurately diagnose perineural invasion in surgically resected samples. At present, the identification of perineural invasion is mainly performed by pathologists manually observing whether there are nerves on the entire pathological section under a microscope and whether there are tumor cells invading these nerves. The high-intensity identification and evaluation of a large number of samples will increase a lot of time and manpower burden, inevitably hindering the speed of the diagnostic process and may affect its accuracy. Therefore, these shortcomings often lead to delayed reporting and a relatively high rate of missed diagnosis. In theory, accurate assessment of PNI requires meticulous evaluation by a multidisciplinary team consisting of biological oncologists, pathologists, surgeons, and radiologists. The use of specialized staining techniques such as immunohistochemistry or immunofluorescence staining will be of great help. All of these have greatly increased the workload. In recent years, great progress has been made in the field of medical intelligent detection, especially in the field of tumor detection and classification. Traditional methods for identifying tumor cells and environments rely heavily on manual detection and subjective judgment, which is time-consuming and error-prone. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a neural tissue recognition method based on artificial intelligence, which solves the problem that traditional tumor cell and environment recognition methods rely heavily on manual detection and subjective judgment, which is time-consuming and prone to errors.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a neural tissue recognition method based on artificial intelligence, comprising the following steps:

[0007] S1. Data Collection

[0008] HE pathological whole-slice images of tumor patients were selected as data, and the tumor types included bladder cancer, pancreatic cancer, and hilar cholangiocarcinoma;

[0009] S2. Data preparation

[0010] All data were annotated using the QuPath full-slice image annotation tool to obtain mask images of positive areas. After obtaining the mask images of the positive areas, a 1024×1024 pixel window was used to extract the positive areas, and the coordinate information was recorded to facilitate cutting of the corresponding areas on the pathological sections. 70% of each type of image was randomly selected as a training set, 10% as a validation set, and 20% as an independent test set, and sections containing neural tissue areas were extracted.

[0011] S3. Model construction

[0012] nnUnet was used to develop a segmentation model based on HE whole-slice image data, integrating feature extraction module and feature fusion module;

[0013] S4. Model training

[0014] In each experiment, the model was trained for 150 cycles, and the initial learning rate was set to 0.01 and gradually decreased during the training process. The Nesterov momentum algorithm was also used to better adjust the learning rate. The sum of the cross entropy loss and the Dice loss was selected as the objective function. Six TITAN V GPUs were used for training, with a batch size of 2 for each GPU. The Dice coefficient was selected as the evaluation metric.

[0015] S5. Active Learning

[0016] An active learning method was used. After each training, all training data sets were inferred, and the areas predicted to be positive but not marked by doctors were resampled as error sets. Each time, 1,000 slices were randomly selected from the error set and incorporated into the training data set for further training. After three iterations, the performance of the model stabilized.

[0017] S6. Model Inference

[0018] After obtaining the trained model, infer the test set data;

[0019] S7. Model performance statistics

[0020] The predicted probability of each pixel is 0.5 as the threshold. If it is greater than 0.5, it is predicted to be positive; if it is less than 0.5, it is predicted to be negative. The Dice coefficient is used to evaluate the accuracy of local area prediction:

[0021]

[0022] Among them, X represents the predicted probability map, and Y represents the actual mask map;

[0023] Due to the region recognition and segmentation on the pathology slice image, a large number of false positives will occur. The performance is evaluated at the pathology slice level by checking whether the slice level contains positive regions. The sensitivity and specificity are as follows:

[0024] Sensitivity:

[0025] Specificity:

[0026] Among them, TP represents slices containing true positives, FP represents false positives, FN represents false negatives, and TN represents true negatives.

[0027] Preferably, the feature extraction module follows a standard convolutional network architecture, comprising a repeated sequence of two 3x3 convolutions, each followed by an instance normalization layer, a Leaky ReLU activation layer, and a 2x2 maximum pooling operation with a stride of 2, and the downsampling process is achieved by replicating the number of feature channels at each interval.

[0028] Preferably, each progression of the feature fusion module includes upsampling of the feature map, followed by a 2x2 transposed convolution operation to halve the number of feature channels, and then concatenating it with the corresponding cropped feature map from the feature extraction module. The processed image then undergoes two 3x3 convolutions, each followed by the same normalization and activation layers. In the last layer, a 1x1 convolution is used to map the feature vector to the target category (neural tissue region).

[0029] Preferably, the specific process of model inference in step S6 is as follows:

[0030] First, the pathological slice thumbnails were processed using the Otsu method to generate a mask representing the tissue area. Based on the tissue area mask, a window of 1024×1024 pixels was used to scale at 20 times magnification. Within the window, if the proportion of the tissue area was greater than or equal to 60%, it was considered a valid slice and slid with a step size of 512 pixels to crop the pathological slice. The cropped image was input into the segmentation model to obtain a 1024×1024 inferred mask. Afterwards, the slice masks were spliced ​​together to obtain a complete inferred mask.

[0031] Preferably, when the slice masks are spliced, two adjacent 1024×1024 masks are overlapped and spliced, and the overlapping 512 pixels are composed of two parts: the 256 pixels of the left half use the rightmost 256 pixels of the left mask, and the 256 pixels of the right half use the leftmost 256 pixels of the right mask.

[0032] The present invention provides a neural tissue recognition method based on artificial intelligence, which has the following beneficial effects:

[0033] 1. The present invention uses AI algorithms to display all neural structures in tumor sections, allowing doctors to determine the presence and incidence of perineural infiltration. Through this model, we provide doctors with more comprehensive information, allowing them to make more accurate and comprehensive judgments, thereby greatly improving the speed and accuracy of clinical diagnosis.

[0034] 2. This invention uses algorithms and deep learning techniques to accurately and quickly assess perineural infiltration in pathology, greatly reducing the workload of manual identification and improving the accuracy of diagnosis. Machine identification of perineural invasion of tumors has obvious practicality, importance, clinical diversity and actual clinical value, demonstrating the potential of deep learning technology to improve tumor detection and diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a schematic diagram of the correlation between perineural infiltration and the malignant biological behavior of solid tumors of the present invention;

[0036] Figure 2 It is a schematic diagram of a data set of the present invention;

[0037] Figure 3 A schematic diagram of a corresponding mask label of a slice containing a neural tissue region of the present invention;

[0038] Figure 4 It is a schematic diagram of the segmentation model of the present invention;

[0039] Figure 5 A schematic diagram showing an example visualization of the identification of peripheral nerves in pancreatic cancer according to the present invention;

[0040] Figure 6 A schematic diagram showing an example of visualizing the identification of peripheral nerves in hilar cholangiocarcinoma of the present invention;

[0041] Figure 7 It is a schematic diagram of the visualization result of the bladder cancer cutting image of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Example:

[0044] like Figure 1-7 As shown, an embodiment of the present invention provides a neural tissue recognition method based on artificial intelligence, comprising the following steps:

[0045] S1. Data Collection

[0046] 452 HE pathological whole-slice images from 452 tumor patients were selected as data. The tumor types included bladder cancer, pancreatic cancer, and hilar cholangiocarcinoma, including 210 images of bladder cancer, 100 images of pancreatic cancer, and 142 images of hilar cholangiocarcinoma. Figure 2 As shown in A, 49 pancreatic cancer patients had detailed clinical information, with a median survival time of 16 months, ranging from 2 months to 59 months. All patients participating in this study signed an informed consent form, which included a statement that pathological tissue and clinical data were used for scientific research; when selecting the patient's HE pathological full-section images, the patient needed to have a clear tumor diagnosis confirmed by histopathological examination; and the patient must have high-quality pathological sections that can be used for analysis without tissue autolysis, structural damage, cell degeneration and wrinkling; and the patient must not receive treatment before surgery. At the same time, the patient is required to sign an informed consent form to allow his pathological data to be used for research. When selecting the patient's HE pathological full-section images, first of all, other non-tumor lesions or benign tumors need to be excluded; patients with poor quality pathological sections also need to be excluded; finally, patients who have received treatment that may affect pathological characteristics before sample collection, such as chemotherapy or radiotherapy, are excluded.

[0047] S2. Data preparation

[0048] All data were annotated using the QuPath full-slice image annotation tool to obtain the mask image of the positive area. After obtaining the mask image of the positive area, a 1024×1024 pixel window was used to extract the positive area, and the coordinate information was recorded to facilitate cutting the corresponding area on the pathological slice. The visualization result of the cut image is shown in the figure. Figure 7 As shown; 70% of each type of image is randomly selected as a training set, 10% as a validation set, and 20% as an independent test set, and 4489 slices containing neural tissue areas are extracted, as shown Figure 2 As shown in B, the corresponding mask label is Figure 3 As shown, in the data preparation process, according to Figure 2 The flowchart shown in C is carried out;

[0049] S3. Model construction

[0050] nnUnet was used to develop a segmentation model based on HE full-slice image data, integrating feature extraction module and feature fusion module. The network structure is as follows Figure 4 As shown;

[0051] S4. Model training

[0052] In each experiment, the model was trained for 150 cycles, and the initial learning rate was set to 0.01 and gradually decreased during the training process. The Nesterov momentum algorithm was also used to better adjust the learning rate. The sum of the cross entropy loss and the Dice loss was selected as the objective function. Six TITAN V GPUs were used for training, with a batch size of 2 for each GPU. The Dice coefficient was selected as the evaluation metric.

[0053] S5. Active Learning

[0054] Each WSI contains a large number of pure negative segments. If these samples are not input into the model, the algorithm may have poor specificity. To solve this problem, an active learning method was used. After each training, all training data sets were inferred, and the areas predicted to be positive but not marked by doctors were resampled as error sets. Each time, 1,000 slices were randomly selected from the error set and incorporated into the training data set for further training. After three iterations, the performance of the model stabilized.

[0055] S6. Model Inference

[0056] After obtaining the trained model, infer the test set data;

[0057] S7. Model performance statistics

[0058] The predicted probability of each pixel is 0.5 as the threshold. If it is greater than 0.5, it is predicted to be positive; if it is less than 0.5, it is predicted to be negative. The Dice coefficient is used to evaluate the accuracy of local area prediction:

[0059]

[0060] Among them, X represents the predicted probability map, and Y represents the actual mask map;

[0061] Due to the region recognition and segmentation on the pathology slice image, a large number of false positives will occur. The performance is evaluated at the pathology slice level by checking whether the slice level contains positive regions. The sensitivity and specificity are as follows:

[0062] Sensitivity:

[0063] Specificity:

[0064] Among them, TP represents slices containing true positives, FP represents false positives, FN represents false negatives, and TN represents true negatives.

[0065] The feature extraction module follows the standard convolutional network architecture, consisting of a repeated sequence of two 3x3 convolutions, each followed by an instance normalization layer, a Leaky ReLU activation layer, and a 2x2 maximum pooling operation with a stride of 2. The downsampling process is achieved by replicating the number of feature channels at each interval.

[0066] Each progression of the feature fusion module includes upsampling of the feature map, followed by a 2x2 transposed convolution operation to halve the number of feature channels. Next, it is concatenated with the corresponding cropped feature map from the feature extraction module. The processed image undergoes two 3x3 convolutions, each followed by the same normalization and activation layers. In the last layer, a 1x1 convolution is used to map the feature vector to the target category (neural tissue region).

[0067] In step S6, the specific process of model inference is as follows:

[0068] First, the pathological slice thumbnails were processed using the Otsu method to generate a mask representing the tissue area. Based on the tissue area mask, a window of 1024×1024 pixels was used to scale at 20 times magnification. Within the window, if the proportion of the tissue area was greater than or equal to 60%, it was considered a valid slice and slid with a step size of 512 pixels to crop the pathological slice. The cropped image was input into the segmentation model to obtain a 1024×1024 inferred mask. Afterwards, the slice masks were spliced ​​together to obtain a complete inferred mask.

[0069] When stitching slice masks, two adjacent 1024×1024 masks are overlapped and stitched, and the overlapping 512 pixels are composed of two parts: the 256 pixels of the left half use the rightmost 256 pixels of the left mask, and the 256 pixels of the right half use the leftmost 256 pixels of the right mask.

[0070] For the Dice coefficient, sensitivity, and specificity, 1000 bootstrap samples were generated by randomly sampling 200 slices from the validation and test sets, and the distributions of these bootstrap statistics were used to estimate 95% confidence intervals (CIs). Overall survival was analyzed using the Mantel–Cox log-rank test. A two-sided P value < 0.05 was considered statistically significant.

[0071] In 49 pancreatic cancer patients with detailed clinical information, perineural invasion was found in pathological sections of solid tumors. Figure 1 A, 51.0% (25 cases) with perineural infiltration (see Figure 1 B). The incidence of perineural invasion is higher in patients with lymph node metastases (see Figure 1 C) and is associated with patients with larger tumors (see Figure 1 D), but it did not reach statistical significance. The overall survival time of patients with perineural invasion was worse than that of patients without perineural invasion (P<0.05) (see Figure 1 E).

[0072] QuPath software was used to annotate the neural tissue regions of 452 HE-stained whole-slice images from bladder cancer, pancreatic cancer, and hilar cholangiocarcinoma. 70% of them were used as training sets, 10% as validation sets, and 20% as independent test sets to train the nnUnet model. The entire training process was combined with an active learning strategy to enhance the specificity of the model.

[0073] Since most areas of pathological sections do not contain neural tissue, initial training only segmented areas containing neural tissue. The initial training set contained 3221 sections. The trained model was tested in all pathological sections of the validation set to determine whether each section level contained neural tissue. The sensitivity and specificity of the validation set were 0.922 (95% CI 0.918, 0.926) and 0.321 (95% CI 0.313, 0.330), respectively. For sections that specifically contain neural tissue, the Dice coefficient of pixel-level performance was 0.798 (95% CI 0.789, 0.806). However, the model tends to misidentify a large number of non-neural areas, mainly due to insufficient learning of negative area features during training and algorithmic limitations.

[0074] In addition, an active learning strategy is adopted to enhance the specificity of the model. First, the WSI of the training set is inferred using the trained model. Then, slices are extracted from the regions where all negative samples are misclassified as positive. Next, 1,000 of these slices are randomly selected to be added to the training set, and the model is retrained.

[0075] After two rounds of iterations, the specificity of the model increased from 0.321 (95% CI 0.313, 0.330) to 0.751 (95% CI 0.743, 0.759) with almost no loss in sensitivity. The performance dropped slightly after the third round of iterations, so we used the results of two rounds for subsequent testing. This strategy guides the model to pay more attention to samples that are difficult to accurately classify, thereby improving its performance and accuracy. For comparison, we also used 2,000 pure negative and positive samples randomly sampled initially for training. The specificity of the model was only 0.427 (95% CI 0.419, 0.436) and the sensitivity was 0.893 (95% CI 0.885, 0.901). Therefore, compared with random sampling, the active learning strategy is able to achieve better performance, as shown in Table 1.

[0076] Finally, the performance of the pathological sections of the three tumor types was tested on an independent test set. The results are shown in Table 2. Figure 5 and Figure 6 Provides a comprehensive review of pancreatic cancer and hilar cholangiocarcinoma ( Figure 6 ) visual example of identifying peripheral nerves, Figure 5 Visualize an example of peripheral nerve identification in pancreatic cancer. Figure 6 Visualize an example of the identification of peripheral nerves in hilar cholangiocarcinoma. Performance metrics for each cancer type include sensitivity, specificity, and Dice coefficient. For bladder cancer, the model has a sensitivity of 0.926 (95% CI: 0.921, 0.931), a specificity of 0.777 (95% CI: 0.769, 0.785), and a Dice coefficient of 0.830 (95% CI: 0.821, 0.838). The performance of pancreatic cancer is slightly lower, with a sensitivity of 0.911 (95% CI: 0.905, 0.916), a specificity of 0.722 (95% CI: 0.718, 0.725), and a Dice coefficient of 0.822 (95% CI: 0.817, 0.826). Hilar cholangiocarcinoma performed best among the three, with a sensitivity of 0.930 (95% CI: 0.921, 0.938), a specificity of 0.750 (95% CI: 0.742, 0.758), and a Dice coefficient of 0.865 (95% CI: 0.861, 0.869).

[0077] Overall, the prediction performance of the three cancers was similar, but the performance of pancreatic cancer was slightly weaker. This was mainly due to some small pancreatic duct glandular tissue being misidentified as neural tissue areas.

[0078] Table 1: Validation set test results based on active learning model optimization

[0079]

[0080] Table 2: Performance of different tumor models on independent test sets

[0081]

[0082]

[0083] The precise localization of nerves within the tumor region and the comprehensive assessment of peripheral nerve invasion provide valuable guidance for clinicians. In this invention, a model was constructed to identify nerve tissue by combining the nnUnet model with an active learning strategy, and its performance was verified in three types of cancer. Compared with the random selection strategy alone, the model achieved better performance through training.

[0084] Intelligent algorithms have shown their advantages in reducing the time required for identification and improving diagnostic efficiency. Most importantly, these algorithms eliminate the influence of subjective factors and ensure a high level of objectivity and consistency. Therefore, diagnostic errors and variations associated with human interpretation can be reduced. In the present invention, the algorithm is able to present nerves in tumor slices, providing a valuable tool to help identify perineural infiltration that may be overlooked by doctors. Artificial intelligence has great potential as a general tool to help doctors develop personalized treatment plans. By quickly and accurately detecting and evaluating tumor invasion into nerves, technology can minimize the risks associated with surgery and radiotherapy. Therefore, treatment effects can be improved and the use of medical resources can be optimized, thereby reducing medical costs. Ultimately, these advances will lead to a significant improvement in patients' quality of life and survival rates.

[0085] However, there are still some limitations in the performance of current models, including sensitivity and specificity. Although peripheral nerves share certain common features, the significant heterogeneity of nerve tissue characteristics after tumor invasion, as well as differences between nerves of different sizes and cross-sections, affect the generalization ability of the model and may lead to changes in specificity. In addition, relying on manually annotated training data also introduces variability, which may affect the performance of the algorithm. Annotation errors may lead to misclassification or missed detection. Dividing pathological sections into smaller sections may also lead to the loss of contextual information, especially in large or elongated nerve tissue, which further affects the ability of the model to distinguish between nerve tissue and surrounding structures. In addition, despite the utilization of a large number of pathological sections, the size and diversity of the dataset may still limit the robustness of the model, especially in less common tumor manifestations. Based on this, doctors can quickly identify suspicious areas and perform rapid verification, thereby achieving fast and accurate annotation. In addition, given the unique characteristics of neural tissue, increasing the number of samples for individual cases and expanding the diversity of tumor types are necessary to enhance the generalization ability of all cancer types. Finally, improving the algorithm itself can improve the accuracy of identifying peripheral nerves in various cancers.

[0086] In summary, machine identification of perineural invasion by tumors has clear utility, importance, clinical diversity, and real clinical value. The model will serve as a complementary tool to traditional diagnostic methods rather than a stand-alone solution. This integration can help clinicians make more informed decisions, optimize treatment options, and improve patient outcomes.

[0087] In summary: This paper proposes a deep learning-based neural recognition method for biliary and pancreatic tumors. The algorithm and deep learning technology are used to accurately and quickly evaluate perineural infiltration in pathology, which greatly reduces the workload of manual recognition and improves the accuracy of diagnosis. This study contributes to the field of medical pathology diagnosis and demonstrates the potential of deep learning technology to improve tumor detection and diagnosis.

[0088] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A neural tissue identification method based on artificial intelligence, characterized in that: The following steps are involved: S1. Data Collection HE pathological whole-slice images of tumor patients were selected as data, and the tumor types included bladder cancer, pancreatic cancer, and hilar cholangiocarcinoma; S2. Data preparation All data were annotated using the QuPath whole-slide image annotation tool to obtain mask images of positive areas; After obtaining the mask map of the positive area, a window of 1024×1024 pixels was used to extract the positive area, and the coordinate information was recorded to facilitate cutting the corresponding area on the pathological section. 70% of each type of image was randomly selected as a training set, 10% as a validation set, and 20% as an independent test set to extract the sections containing the neural tissue area. S3. Model construction nnUnet was used to develop a segmentation model based on HE whole-slice image data, integrating feature extraction module and feature fusion module; S4. Model training In each experiment, the model was trained for 150 cycles, and the initial learning rate was set to 0.01 and gradually decreased during the training process. The Nesterov momentum algorithm was also used to better adjust the learning rate. The sum of the cross entropy loss and the Dice loss was selected as the objective function. Six TITAN V GPUs were used for training, with a batch size of 2 for each GPU. The Dice coefficient was selected as the evaluation metric. S5. Active Learning An active learning method was used. After each training, all training data sets were inferred, and the areas predicted to be positive but not marked by doctors were resampled as error sets. Each time, 1,000 slices were randomly selected from the error set and incorporated into the training data set for further training. After three iterations, the performance of the model stabilized. S6. Model Inference After obtaining the trained model, infer the test set data; S7. Model performance statistics The predicted probability of each pixel is 0.5 as the threshold. If it is greater than 0.5, it is predicted to be positive; if it is less than 0.5, it is predicted to be negative. The Dice coefficient is used to evaluate the accuracy of local area prediction: Among them, X represents the predicted probability map, and Y represents the actual mask map; Due to the region recognition and segmentation on the pathology slice image, a large number of false positives will occur. The performance is evaluated at the pathology slice level by checking whether the slice level contains positive regions. The sensitivity and specificity are as follows: Sensitivity: Specificity: Among them, TP represents slices containing true positives, FP represents false positives, FN represents false negatives, and TN represents true negatives.

2. The neural tissue identification method based on artificial intelligence according to claim 1, characterized in that: The feature extraction module follows the standard convolutional network architecture, consisting of a repeated sequence of two 3x3 convolutions, each followed by an instance normalization layer, a Leaky ReLU activation layer, and a 2x2 maximum pooling operation with a stride of 2. The downsampling process is achieved by replicating the number of feature channels at each interval.

3. The neural tissue identification method based on artificial intelligence according to claim 1, characterized in that: Each progression of the feature fusion module includes upsampling of the feature map, followed by a 2x2 transposed convolution operation to halve the number of feature channels. Next, it is concatenated with the corresponding cropped feature map from the feature extraction module. The processed image undergoes two 3x3 convolutions, each followed by the same normalization and activation layers. In the last layer, a 1x1 convolution is used to map the feature vector to the target category (neural tissue region).

4. The neural tissue identification method based on artificial intelligence according to claim 1, characterized in that: The specific process of model inference in step S6 is as follows: First, the pathological slide thumbnails were processed using the Otsu method to generate masks representing tissue regions. Based on the tissue region masks, a window of 1024 × 1024 pixels was used for scaling at 20x magnification. Within the window, if the proportion of the tissue area is greater than or equal to 60%, it is considered a valid slice and is slid with a step size of 512 pixels to crop the pathological slice; The cropped image is fed into the segmentation model to obtain an inferred mask of 1024×1024; afterward, the sliced ​​masks are stitched together to obtain the complete inferred mask.

5. The neural tissue identification method based on artificial intelligence according to claim 4, characterized in that: When stitching slice masks, two adjacent 1024×1024 masks are overlapped and stitched, and the overlapping 512 pixels are composed of two parts: the 256 pixels of the left half use the rightmost 256 pixels of the left mask, and the 256 pixels of the right half use the leftmost 256 pixels of the right mask.

Images