Deep learning-based prediction method and system for invasion and metastasis of adenoid cystic carcinoma

Abstract

The invention provides a deep learning-based prediction method and system for invasion and metastasis of adenoid cystic carcinoma. Specifically, an adenoid cystic carcinoma invasion classification model and an adenoid cystic carcinoma invasion mapping model are constructed based on adenoid cystic carcinoma full-width slice images (WSIs). According to the adenoid cystic carcinoma invasion classification and mapping model, invasion classification of a certain adenoid cystic carcinoma pathological tissue to be detected can be accurately and efficiently evaluated, and an invasion position mark is given.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital pathological image processing and decision support, and specifically, relates to a method and system for predicting the invasion and metastasis of adenoid cystic carcinoma based on deep learning. Background Art

[0002] Adenoid cystic carcinoma (AdCC) is a rare and highly aggressive cancer. Commonly reported cases include the lacrimal gland, breast, cervix, esophagus, lung, and prostate. The clinical diagnosis of AdCC relies heavily on histopathological examination, which is considered the "gold standard", especially when the treatment plan includes extensive surgery, which may even involve sacrificing the facial nerve.

[0003] Histopathological analysis of whole-slice images (WSIs) has become a critical part of the diagnostic workflow. However, traditional methods for evaluating WSIs can be time-consuming and subject to interobserver variability, highlighting the need for more efficient and reliable diagnostic methods.

[0004] Therefore, there is an urgent need to develop a method that can quickly and accurately classify and identify the invasion area based on WSIs, which can provide effective assistance for the diagnosis and treatment strategies of AdCC. Summary of the invention

[0005] The present invention provides a method capable of quickly and accurately classifying and identifying the invasion area.

[0006] In a first aspect of the present invention, a method for constructing an invasion and metastasis classification model of adenoid cystic carcinoma (AdCC) based on deep learning is provided, comprising the steps of:

[0007] (s1) providing a first data set, wherein the first data set comprises WSI images of adenoid cystic carcinoma pathological tissues, and the images are WSI images annotated with invasion;

[0008] (s2) constructing a classification model for invasion and metastasis of adenoid cystic carcinoma based on the first data set; and

[0009] (s3) constructing an adenoid cystic carcinoma invasion mapping model based on the first data set.

[0010] In another preferred embodiment, in step (s2), the following steps are included:

[0011] The first data set is divided into a training set and a test set according to a certain ratio (training set: test set = 0.85:0.15); in the training set, a five-fold cross-validation is used to find the optimal parameters of the model; then the optimal model is constructed based on the optimal parameters and verified on the test set.

[0012] In another preferred embodiment, in step (s3), the following steps are included:

[0013] The first data set is divided into a training set and a test set according to a certain ratio (training set: test set = 0.85:0.15); in the training set, a five-fold cross-validation is used to find the optimal parameters of the model; then the optimal model is constructed based on the optimal parameters and verified on the test set.

[0014] In another preferred example, the adenoid cystic carcinoma invasion mapping model is obtained by training the LinkNet decoder after the adenoid cystic carcinoma invasion and metastasis classification model is trained.

[0015] In another preferred embodiment, step (s2) further includes the following sub-steps:

[0016] (s2a) Image preprocessing and enhancement: performing color homogenization on the images in the first data set, and dividing the homogenized WSI images into n small blocks, where n≥40 and n≤60, by high-pass filtering and threshold limiting, so as to obtain image blocks containing tumor areas; and performing data enhancement on each of the image blocks;

[0017] (s2b) Feature extraction: using a SEResNet50 encoder to extract image features from the image block after data enhancement, thereby obtaining low-dimensional image features of the image block;

[0018] (s2c) Invasion and metastasis classification: The low-dimensional image features are input into the fully connected layer for processing to obtain the tumor invasion and metastasis classification of the image; and the image blocks are integrated into the original image using the majority agreement principle to obtain the WSI image invasion and metastasis classification.

[0019] In another preferred embodiment, n≥45 and n≤55.

[0020] In another preferred embodiment, the value of n is any of the following positive integers: 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55.

[0021] In another preferred embodiment, in step (s3), sub-steps (s2a), (s2b) and the following sub-steps are also included:

[0022] (s3a) Invasion mapping: The low-dimensional image features are input into the LinkNet decoder to obtain the tumor invasion location annotation of the image block.

[0023] In another preferred embodiment, in step (s2a), the image is subjected to color uniformization processing using the Reinhard method.

[0024] In another preferred example, in step (s2a), the data enhancement method is selected from the following group: random rotation, random flipping, random scaling, random gamma adjustment, or a combination thereof.

[0025] In another preferred embodiment, the majority agreement principle is that when most of the picture blocks show the same invasion and transfer classification, it is determined that the original image (or WSI image) after the picture blocks are integrated also has the invasion classification.

[0026] In the present invention, the majority agreement principle refers to voting using the prediction results of 50 patches in the same whole slice image (WSI), and finally taking the result with the most votes as the final prediction. The prediction result of each patch is equivalent to a "vote", and the final prediction follows the principle of "the minority obeys the majority". By integrating the judgments of multiple patches and summarizing local information, the prediction accuracy of the entire WSI can be improved.

[0027] In another preferred embodiment, the invasion and metastasis classification is selected from the following group: perineural invasion, vascular invasion, lymph node metastasis, or a combination thereof.

[0028] In another preferred embodiment, the invasion map is selected from the following group: peripheral nerve invasion map, vascular invasion map.

[0029] In a second aspect of the present invention, a deep learning-based adenoid cystic carcinoma (AdCC) invasion and metastasis classification system is provided, comprising:

[0030] (M1) an input module, the input module being configured to input a WSI image of adenoid cystic carcinoma pathological tissue to be detected;

[0031] (M2) a classification module, wherein the classification module is configured to classify the WSI image of the adenoid cystic carcinoma pathological tissue to be tested based on an adenoid cystic carcinoma invasion-metastasis classification model; wherein the adenoid cystic carcinoma invasion-metastasis classification model is an adenoid cystic carcinoma invasion-metastasis classification model constructed by a deep learning method based on the WSI image of the adenoid cystic carcinoma pathological tissue annotated with invasion;

[0032] Alternatively, the adenoid cystic carcinoma invasion and metastasis classification model is an adenoid cystic carcinoma invasion and metastasis classification model constructed by the construction method described in the first aspect of the present invention; and

[0033] (M3) Output module: The output module is configured to output the classification result.

[0034] In another preferred embodiment, the classification system further includes (M2a) a mapping module, wherein the mapping module is configured to map and annotate the WSI image of the adenoid cystic carcinoma pathological tissue to be tested based on the adenoid cystic carcinoma invasion mapping model; wherein the adenoid cystic carcinoma invasion mapping model is an adenoid cystic carcinoma invasion mapping model constructed by a deep learning method based on the WSI image of the adenoid cystic carcinoma pathological tissue annotated with invasion;

[0035] Alternatively, the adenoid cystic carcinoma invasion mapping model is an adenoid cystic carcinoma invasion mapping model constructed by the construction method described in the first aspect of the present invention.

[0036] In another preferred example, the output module is further configured to output the invasion mapping result.

[0037] In another preferred example, the system further includes (M4) a control module: the control module is configured to control the operation of each module.

[0038] In a third aspect of the present invention, a method for evaluating the invasion and metastasis classification of adenoid cystic carcinoma pathological tissue to be tested is provided, comprising the steps of:

[0039] (a) providing a WSI image of adenoid cystic carcinoma pathological tissue to be tested;

[0040] (b) inputting the WSI image of the adenoid cystic carcinoma pathological tissue to be tested into the adenoid cystic carcinoma invasion and metastasis classification model and the adenoid cystic carcinoma invasion mapping model using the method described in the first aspect of the present invention;

[0041] (c) evaluating the WSI image of the adenoid cystic carcinoma pathological tissue to be tested using the adenoid cystic carcinoma invasion and metastasis classification model, and optionally the adenoid cystic carcinoma invasion mapping model, thereby obtaining a tumor invasion and metastasis classification and invasion mapping of the WSI image of the adenoid cystic carcinoma pathological tissue to be tested; and

[0042] (d) Outputting the evaluation result.

[0043] In another preferred embodiment, the method is an in vitro method.

[0044] In another preferred embodiment, the method is non-diagnostic and non-therapeutic.

[0045] In another preferred embodiment, the invasion and metastasis classification is selected from the following group: perineural invasion, vascular invasion, lymph node metastasis, or a combination thereof.

[0046] In another preferred embodiment, the invasion map is selected from the following group: peripheral nerve invasion map, vascular invasion map.

[0047] It should be understood that within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features specifically described below (such as embodiments) can be combined with each other to form a new or preferred technical solution. Due to space limitations, they will not be described one by one here. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The distribution of cancer lesions in the patient's oral anatomy is shown.

[0049] Figure 2 An overview of the deep learning based model in the present invention is shown.

[0050] Figure 3 ROC curves comparing subgraph level and sliding level analysis by the model of the present invention are shown.

[0051] Figure 4 Examples of tumor infiltration are shown (red - neural invasion, yellow - vascular invasion). DETAILED DESCRIPTION

[0052] After extensive and in-depth research, the inventors have developed for the first time a highly accurate and precise adenoid cystic carcinoma invasion and metastasis model, as well as an invasion mapping model constructed on this basis. Specifically, through adenoid cystic carcinoma full-section images (WSI), a deep learning method was used to construct an adenoid cystic carcinoma invasion and metastasis classification model and an invasion mapping model; wherein the invasion and metastasis include: perineural invasion, vascular invasion, and / or lymph node metastasis, and the mapping includes: peripheral nerve invasion mapping and / or vascular invasion mapping. The present invention was completed on this basis.

[0053] the term

[0054] In order to more easily understand the present disclosure, some terms are first defined. As used in this application, unless otherwise expressly provided herein, each of the following terms should have the meaning given below. Other definitions are set forth throughout the application.

[0055] The term "about" can refer to a value or composition that is within an acceptable error range for a particular value or composition as determined by one of ordinary skill in the art, which will depend in part on how the value or composition is measured or determined. For example, as used herein, the expression "about 100" includes all values ​​between 99 and 101 (e.g., 99.1, 99.2, 99.3, 99.4, etc.).

[0056] As used herein, the term "comprising" or "including (comprising)" may be open, semi-closed and closed. In other words, the term also includes "consisting essentially of" or "consisting of".

[0057] As used herein, unless otherwise indicated, any concentration range, percentage range, ratio range or integer range should be understood to include values ​​of any integer within the range and fractional values ​​thereof (e.g., one tenth and one hundredth of an integer) where appropriate.

[0058] As used herein, the term "and / or" refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0059] As used in this article, "infiltration", "invasion" and "aggression" are used interchangeably and refer to the process in which tumor cells damage the original and surrounding normal tissue structures through various possible pathways and methods, and eventually break away from the primary tumor and disperse in the surrounding tissues to varying degrees.

[0060] As used in this article, the "Dice coefficient", also known as the Dice Similarity Coefficient, is a set similarity measurement function, which is usually used to calculate the similarity or overlap between two samples. The value of the Dice coefficient ranges from 0 to 1. The closer its value is to 1, the higher the overlap between the predicted result and the true label, and the higher the similarity; conversely, the closer its value is to 0, the lower the overlap between the predicted result and the true label, and the lower the similarity.

[0061] Adenoid cystic carcinoma

[0062] Adenoid cystic carcinoma (AdCC) is a rare cancer that primarily originates in the salivary glands. It accounts for approximately 1% of all malignancies in the oral and maxillofacial regions and approximately 22% of all major and minor salivary gland malignancies. Although AdCC most commonly arises in the oral salivary glands, it can occur in almost any site that contains a secretory gland component. Cases that have been reported in the literature include the lacrimal gland, breast, cervix, esophagus, lung, and prostate. AdCC is known for its aggressive behavior, often leading to local invasion and distant metastasis, which can result in a poor prognosis or even fatal outcome.

[0063] Whole Slice Image (WSI)

[0064] A whole slide image (WSI) is a complete digital image generated by digitally scanning a tissue slice through a high-resolution microscope. It covers the entire field of view of the tissue slice and is usually used for pathological analysis, enabling the microscopic structure of the tissue to be viewed at different magnifications. The whole slide image (WSI) used in this project was magnified 40 times. The average size of the WSI is 264,752x121,293 pixels, and the maximum size is 381,861x 174,511 pixels.

[0065] System of the present invention

[0066] The present invention provides a deep learning-based adenoid cystic carcinoma (AdCC) invasion and metastasis classification system, comprising:

[0067] (M1) an input module, the input module being configured to input a WSI image of adenoid cystic carcinoma pathological tissue to be detected;

[0068] (M2) a classification module, wherein the classification module is configured to classify the WSI image of the adenoid cystic carcinoma pathological tissue to be tested based on an adenoid cystic carcinoma invasion-metastasis classification model; wherein the adenoid cystic carcinoma invasion-metastasis classification model is an adenoid cystic carcinoma invasion-metastasis classification model constructed by a deep learning method based on the WSI image of the adenoid cystic carcinoma pathological tissue annotated with invasion;

[0069] Alternatively, the adenoid cystic carcinoma invasion and metastasis classification model is an adenoid cystic carcinoma invasion and metastasis classification model constructed by the construction method described in the first aspect of the present invention; and

[0070] (M3) Output module: The output module is configured to output the classification result.

[0071] In another preferred embodiment, the classification system further includes (M2a) a mapping module, wherein the mapping module is configured to map and annotate the WSI image of the adenoid cystic carcinoma pathological tissue to be tested based on the adenoid cystic carcinoma invasion mapping model; wherein the adenoid cystic carcinoma invasion mapping model is an adenoid cystic carcinoma invasion mapping model constructed by a deep learning method based on the WSI image of the adenoid cystic carcinoma pathological tissue annotated with invasion;

[0072] Alternatively, the adenoid cystic carcinoma invasion mapping model is an adenoid cystic carcinoma invasion mapping model constructed by the construction method described in the first aspect of the present invention.

[0073] In another preferred example, the output module is further configured to output the invasion mapping result.

[0074] In another preferred example, the system further includes (M4) a control module: the control module is configured to control the operation of each module.

[0075] The main advantages of the present invention include:

[0076] (a) The method of the present invention can identify the invaded area rapidly with high sensitivity and specificity.

[0077] (b) The present invention provides visualization results of tumor invasion.

[0078] (c) The method of the present invention can quickly and accurately predict lymph node metastasis.

[0079] The present invention will be further described below in conjunction with specific examples. It should be understood that these examples are intended to illustrate the present invention only and are not intended to limit the scope of the present invention. The experimental methods in the following examples where specific conditions are not specified are usually performed under conventional conditions, such as those described in Sambrook et al., Molecular Cloning: A Laboratory Manual (New York: Cold Spring Harbor Laboratory Press, 1989), or under conditions recommended by the manufacturer. Unless otherwise indicated, percentages and parts are weight percentages and weight parts.

[0080] Example 1

[0081] 1.1 Materials and methods

[0082] 1.1.1 Whole-slice image acquisition

[0083] Methods: A total of 958 whole-section images (WSIs) of adenoid cystic carcinoma of the head and neck collected from the Department of Oral Pathology, the Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine were obtained and analyzed. The specimens obtained were formalin-fixed, paraffin-embedded tissues, and stained with hematoxylin and eosin (H&E). The slides were digitized using a Panoramic MIDI Scanner (3DHISTECH, Budapest, Hungary) at 40x eyepiece magnification.

[0084] The results are as follows Figure 1 As shown. Figure 1 The distribution of cancer lesions in the patients' oral anatomy was known. The most common site was the palate (n = 244), followed by the parotid gland (n = 111) and the tongue (n = 105). The submandibular region and floor of mouth were found in 104 and 98 cases, respectively. Other sites included the maxilla (n = 60), sublingual gland (n = 56), and buccal region (n = 53), while the mandible (n = 33), maxillary sinus (n = 28), and gingiva (n = 20) were less common.

[0085] The diagnosis was established by expert pathological review, immunohistochemistry, and clinical data. The average size of the whole-section image (WSI) was 264,752 × 121,293 pixels, and the maximum size was 381,861 × 174,511 pixels. This study was approved by the Institutional Ethics Committee of Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine.

[0086] The study cohort included 958 patients with an average age of 52.25±14.08 years, including 452 males and 506 females. The incidence of perineural invasion was 63.78%, the incidence of vascular invasion was 17.33%, and the incidence of lymph node metastasis was 8.25%, as shown in Table 1. Among them, the patients were divided into eight groups according to the type and degree of invasion.

[0087] Table 1

[0088]

[0089] 1.1.2 Sub-image acquisition and processing

[0090] In view of the fact that the data of whole slice images (WSIs) is too large to be directly processed by deep learning models, the present invention adopts a sub-graph-based method to train and test the model. That is, the present invention adopts a method of using sub-graphs (patches) for model training and testing. Specifically, it divides the original full-width image into smaller parts, namely sub-graphs. During the model training and testing process, the model learns and predicts each sub-graph separately, so that the model can process complex or large-scale data more effectively, and finally summarizes the results of each sub-graph to obtain an accurate prediction of the overall image or data.

[0091] This approach partitions WSIs into smaller, manageable subgraphs for model input, and then uses subgraph-to-slice techniques to aggregate model predictions from these subgraphs back to the slice level for easier decision making.

[0092] The initial step of the process of the present invention involves color standardization of the image using the Reinhard method. This standardization is essential to maintain the consistency of the appearance of different slice images. Subsequently, the present invention mainly extracts sub-images from the tumor area in the WSIs. The present invention uses high-pass filtering combined with thresholding techniques to ensure that relevant areas are selected.

[0093] High-pass filtering was applied to highlight features of interest by suppressing low-frequency background information, thereby enhancing the contrast of the tumor region. Thresholding techniques were then used to identify and extract subimages based on predefined intensity criteria, ensuring that only regions with important pathological information were included. The present invention limited the extraction to a maximum of 50 subimages per WSI, each with a size of 512×512 pixels and 40 times magnification. This approach is designed to facilitate detailed analysis and effectively manage the amount of data. A total of 47,900 subimages were collected for downstream tasks.

[0094] To further improve the reliability of the dataset, regions showing peripheral nerve and vascular invasion were annotated by a senior pathologist.

[0095] 1.1.3 Invasion Classification and Mapping Based on Deep Learning

[0096] 958 whole slice images (WSIs) were randomly divided into training and test sets to develop and validate the model of the present invention. Due to the large amount of data in WSIs, they were cropped into smaller sub-images for neural network processing.

[0097] The training set includes 813 slices and 40,650 subgraphs, while the test set includes 145 slices and 7,250 subgraphs. In order to ensure the robustness of the model construction, the present invention adopts a 5-fold cross-validation method. In each fold, the data set is divided into 622 slices (31,100 subgraphs) for training and 191 slices (9,550 subgraphs) for verification.

[0098] In order to prevent data leakage, that is, subgraphs of the same patient appear in the training set or the test set at the same time, the present invention divides the data based on the patient ID, ensuring that all subgraphs belonging to a single patient appear in only one data set.

[0099] The deep learning model detection schematic diagram of the present invention is as follows Figure 2 The output results include two aspects: invasion classification and mapping.

[0100] The model architecture is based on SEResNet50 for feature extraction. Two processing flows were then established.

[0101] The first pipeline focuses on invasion classification at the sub-image level, using latent space features as input, followed by dense layers. Considering the different tumor invasion patterns identified in each sub-image, the sub-image level classification is aggregated to the slice level using a majority voting method.

[0102] The second pipeline uses latent space features to map intrusion-related regions within the input subgraphs of the decoder based on the LinkNet pre-trained network.

[0103] Initially, the model is trained only on the invasion classification class. Once this training is stable, the weights are frozen and the invasion mapping pipeline begins. This sequential approach ensures a solid foundation for the model, first identifying the presence of invasion and then localizing it. In addition, the input sub-maps are dynamically pre-processed and augmented to increase the robustness of the model.

[0104] The enhancement techniques used in the present invention include random rotation, random flipping, random scaling and random gamma adjustment. The performance of the model is evaluated using multiple indicators: accuracy, sensitivity, specificity and area under the receiver operating characteristic curve (AUROC) of invasion classification and Dice coefficient of invasion mapping. The calculation is performed by an NVIDIA GeForce RTX 3090Ti graphics processor.

[0105] 1.2 Results

[0106] 1.2.1 Model Performance for Intrusion Classification

[0107] The study evaluated the model's ability to identify three key pathological features in cancer progression: peripheral nerve invasion, vascular invasion, and lymph node metastasis. The model's performance was evaluated using a test set dataset containing 7,250 subgraphs from 145 patients.

[0108] Table 2 summarizes the performance metrics for sub-image level and slice level classification. At the sub-image level, the model showed high accuracy in identifying the three features. AUROC of 0.93 for nerve invasion, 0.85 for vascular invasion, and 0.85 for lymph node metastasis was observed. The accuracy, sensitivity, and specificity for peripheral nerve invasion were recorded at 0.93, 0.92, and 0.93, respectively. The accuracy for vascular invasion was 0.85, the sensitivity was 0.84, and the specificity was 0.87. The accuracy for lymph node metastasis classification was 0.94, the sensitivity was 0.92, and the specificity was 0.96.

[0109] After using the subgraph-sliding technique, the performance of the model was significantly improved, with the AUROC values ​​of peripheral nerve invasion, vascular invasion, and lymph node metastasis increasing to 0.94, 0.93, and 0.94, respectively. Figure 3 AUROC curves at the sub-image level and slice level are shown. At the slice level, the accuracy of identifying peripheral nerve invasion was 0.94, sensitivity was 0.91, and specificity was 0.98. The accuracy of vascular invasion classification was 0.93, sensitivity was 0.90, and specificity was 0.96. Similarly, for lymph node metastasis, the model achieved an accuracy of 0.95, sensitivity of 0.91, and specificity of 0.98.

[0110] Table 2

[0111]

[0112] 1.2.2 Model Performance of Invasion Mapping

[0113] Evaluation of invasion mapping performance Figure 4 As shown, it is shown that the model is able to identify histopathological features in a complex environment containing multiple benign cells. In addition, the method of the present invention successfully detected the peripheral nerve and blood vessel invasion of tumor cells.

[0114] Therefore, the invasion mapping accurately identified the correct location, although some mapped areas showed obvious gaps. We used the Dice coefficient to measure the similarity of two sets. The Dice coefficient is often used in image segmentation tasks, especially in medical image processing. This is the formula of the Dice coefficient, Dice(A,B)=2(A∩B) / (A+B), where A is the predicted segmented area, B is the true segmented area, and A∩B represents the intersection of the predicted area and the true area (i.e., the area of ​​the overlapping part). Quantitative analysis As shown in Table 3, the Dice coefficient for peripheral nerve invasion is 0.71, indicating the consistency between the model prediction and the pathological results. In the case of vascular invasion, the Dice coefficient is higher at 0.77, showing stronger consistency and indicating that the model is more effective in accurately mapping vascular invasion than peripheral nerve invasion.

[0115] Table 3 Invasion mapping performance on the test set

[0116] Invasion Mapping Dice coefficient Peripheral nerve invasion 0.71 Vascular invasion 0.77

[0117] discuss

[0118] Accurate diagnosis of adenoid cystic carcinoma (AdCC) invasion is a major challenge. Histopathological images help identify invasion, however cases can be missed during manual inspection, especially when cancer cells invade small neural territories. Given the critical importance of this highly aggressive cancer in patient management and further subtyping of malignant cases, there is little room for diagnostic error.

[0119] In this context, we developed the first neural network designed to diagnose AdCC invasion and metastasis, showing successful performance (AUC of 0.94 for peripheral nerve invasion, 0.93 for vascular invasion, and 0.94 for lymph node metastasis). Inspired by research showing that image-level classification can assist segmentation tasks, our model was designed in two stages. A SEResNet50 encoder benefits from training and testing on invasion classification, and then maps the invasion regions. This strategic sequence enhances invasion mapping, successfully identifying peripheral nerve invasion (Dice coefficient of 0.71) and vascular invasion (Dice coefficient of 0.77).

[0120] Existing studies have trained neural networks for AdCC that have focused on differentiating AdCC from basal cell adenomas, but have not yet been used to diagnose invasion and metastasis or identify specific areas of perineural and vascular invasion. The scarcity of AdCC makes it difficult to develop a large number of test samples due to limited sample availability.

[0121] The present invention aims to achieve highly accurate classification by training a neural network using data precisely annotated manually by professional doctors. Although the annotation process is time-consuming and labor-intensive, and relies on the expertise of doctors, resulting in a limited training set size, the invasive areas only account for a small part of the total area of ​​the slice, and accurately locating these areas is crucial for accurate prediction of invasion.

[0122] The method of the present invention can be used for auxiliary diagnosis. The validation results of the present invention show that neural networks can make predictions with clinically useful accuracy. It is worth noting that the model of the present invention accurately predicts lymph node metastasis, which is not directly visible in histological images, probably because the model utilizes cell morphology and identifies peripheral nerve and vascular invasion.

[0123] Distinctive cell morphological features, such as changes in size, shape, and nuclear characteristics, are important indicators of malignancy and reveal the aggressive nature of cancer cells and their potential to invade and metastasize. The severity of perineural and vascular invasion observed may correlate with the likelihood of metastasis, indicating the ability of cancer cells to migrate along neural pathways and enter the circulatory system, which is critical for metastasis to lymph nodes. The diagnosis of lymph node metastasis has important implications for treatment planning, including neck lymph node dissection, postoperative treatment, and follow-up planning.

[0124] In summary, the present invention has developed the first deep learning-based method for invasiveness and metastasis analysis in full-length slice images of adenoid cystic carcinoma (AdCC). The model of the present invention exhibits robust performance and shows potential as an auxiliary diagnostic tool, which is very valuable for diagnosing AdCC. The present invention not only helps to identify this highly invasive and difficult to diagnose malignant tumor, but also highlights the key histological features required for accurate diagnosis. By providing an automated AdCC invasion and metastasis analysis method, new ideas are provided for enhancing patient management and treatment strategies.

[0125] All documents mentioned in the present invention are cited as references in this application, just as each document is cited as reference individually. In addition, it should be understood that after reading the above teachings of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the claims attached to this application.

Claims

1. A method for constructing a classification model for invasion and metastasis of adenoid cystic carcinoma (AdCC) based on deep learning, characterized in that: Includes steps: (s1) providing a first data set, wherein the first data set comprises WSI images of adenoid cystic carcinoma pathological tissues, and the images are WSI images annotated with invasion; (s2) constructing a classification model for invasion and metastasis of adenoid cystic carcinoma based on the first data set; and (s3) constructing an adenoid cystic carcinoma invasion mapping model based on the first data set.

2. The method according to claim 1, characterized in that In step (s2), the following steps are included: The first data set is divided into a training set and a test set according to a certain ratio (training set: test set = 0.85:0.15); in the training set, a five-fold cross-validation is used to find the optimal parameters of the model; then the optimal model is constructed based on the optimal parameters and verified on the test set.

3. The method according to claim 1, characterized in that The adenoid cystic carcinoma invasion mapping model is obtained by training the LinkNet decoder after the adenoid cystic carcinoma invasion and metastasis classification model is trained.

4. The method according to claim 1, characterized in that Step (s2) also includes the following sub-steps: (s2a) Image preprocessing and enhancement: performing color homogenization on the images in the first data set, and dividing the homogenized WSI images into n small blocks, where n≥40 and n≤60, by high-pass filtering and threshold limiting, so as to obtain image blocks containing tumor areas; and performing data enhancement on each of the image blocks; (s2b) Feature extraction: using a SEResNet50 encoder to perform image feature extraction on the image block after data enhancement, thereby obtaining low-dimensional image features of the image block; (s2c) Invasion and metastasis classification: inputting the low-dimensional image features into a fully connected layer for processing, thereby obtaining a tumor invasion and metastasis classification of the image; And by using the majority agreement principle, the image blocks are integrated into the original image to obtain the invasion and metastasis classification of the WSI image.

5. The method according to claim 1, characterized in that In step (s3), sub-steps (s2a), (s2b) and the following sub-steps are also included: (s3a) Invasion mapping: The low-dimensional image features are input into the LinkNet decoder to obtain the tumor invasion location annotation of the image block.

6. The method according to claim 1, characterized in that In step (s2a), the data enhancement method is selected from the following group: random rotation, random flipping, random scaling, random gamma adjustment, or a combination thereof.

7. A deep learning-based adenoid cystic carcinoma (AdCC) invasion and metastasis classification system, characterized in that: include: (M1) an input module, the input module being configured to input a WSI image of adenoid cystic carcinoma pathological tissue to be detected; (M2) a classification module, wherein the classification module is configured to classify the WSI image of the adenoid cystic carcinoma pathological tissue to be tested based on an adenoid cystic carcinoma invasion-metastasis classification model; wherein the adenoid cystic carcinoma invasion-metastasis classification model is an adenoid cystic carcinoma invasion-metastasis classification model constructed by a deep learning method based on the WSI image of the adenoid cystic carcinoma pathological tissue annotated with invasion; Alternatively, the adenoid cystic carcinoma invasion and metastasis classification model is an adenoid cystic carcinoma invasion and metastasis classification model constructed by the construction method according to claim 1; and (M3) Output module: The output module is configured to output the classification result.

8. The classification system according to claim 7, characterized in that The classification system further includes (M2a) a mapping module, wherein the mapping module is configured to map and annotate the WSI image of the adenoid cystic carcinoma pathological tissue to be tested based on the adenoid cystic carcinoma invasion mapping model; wherein the adenoid cystic carcinoma invasion mapping model is an adenoid cystic carcinoma invasion mapping model constructed by a deep learning method based on the WSI image of the adenoid cystic carcinoma pathological tissue with invasion annotation; Alternatively, the adenoid cystic carcinoma invasion mapping model is an adenoid cystic carcinoma invasion mapping model constructed by the construction method according to claim 1.

9. A method for evaluating the invasion and metastasis classification of adenoid cystic carcinoma pathological tissue to be tested, characterized in that: Includes steps: (a) providing a WSI image of adenoid cystic carcinoma pathological tissue to be tested; (b) inputting the WSI image of the adenoid cystic carcinoma pathological tissue to be tested into the method of claim 1 to construct the adenoid cystic carcinoma invasion and metastasis classification model and the adenoid cystic carcinoma invasion mapping model; (c) evaluating the WSI image of the adenoid cystic carcinoma pathological tissue to be tested using the adenoid cystic carcinoma invasion and metastasis classification model, and optionally the adenoid cystic carcinoma invasion mapping model, thereby obtaining a tumor invasion and metastasis classification and invasion mapping of the WSI image of the adenoid cystic carcinoma pathological tissue to be tested; and (d) Outputting the evaluation result.

10. The method according to claim 9, characterized in that The invasion and metastasis classification is selected from the group consisting of perineural invasion, vascular invasion, lymph node metastasis, or a combination thereof; The invasion map is selected from the group consisting of a peripheral nerve invasion map, a vascular invasion map.

Images