IONet model for pathological typing of ovarian cancer based on digital pathological image

Abstract

The invention relates to the technical field of pathological typing, and particularly discloses an IONet model for pathological typing of ovarian cancer based on a digital pathological image, which is technically characterized in that a method for pathological typing of ovarian cancer through the IONet model comprises the following steps: S1, automatic background filtering; s2, multi-scale feature extraction; s3, carrying out the classification of a Transform encoder and an MIL (Maximum Information Language); and S4, an integration strategy based on a reasoning stage. According to the method, the dependence on manual annotation is reduced through automatic background filtering and multi-scale feature extraction technologies, and large-scale WSI processing is promoted. Besides, by utilizing the self-supervised learning ability of DINO-ViT, the IONet model can extract meaningful features under the condition of not depending on artificial tags, and the model development and training process is simplified.

Description

Technical Field

[0001] The present invention relates to the technical field of pathological typing, and in particular to an IONet model for pathological typing of ovarian cancer based on digital pathological images. Background Art

[0002] Ovarian cancer is a malignant tumor that occurs in ovarian tissue. Its early symptoms are not obvious and it is usually discovered in the late stage. The types of ovarian cancer are mainly divided into the following categories: Epithelial ovarian cancer: the most common, accounting for about 90% of ovarian cancer, including serous carcinoma, endometrioid carcinoma, clear cell carcinoma, mucinous carcinoma, etc. Germ cell tumors: less common, usually occurring in young women, such as dysgerminoma, yolk sac tumor, etc. Sex cord-stromal tumors: rare, such as granulosa cell tumor, Sertoli-Leydig cell tumor, etc. Metastatic tumors: cancers in other parts of the body metastasize to the ovaries, such as Krukenberg tumors. Ovarian cancer is a complex and diverse disease. Early diagnosis and treatment are crucial. Through comprehensive treatment and personalized management, the survival rate and quality of life of patients can be improved.

[0003] Clinically, the diagnosis of ovarian cancer usually includes imaging examinations (such as ultrasound, CT, MRI, etc.), blood tests (detection of tumor marker CA-125) and pathological examinations (obtaining tissue samples through biopsy or surgery for pathological analysis).

[0004] In the existing technology, traditional ovarian cancer pathological typing mainly relies on morphological observation, immunohistochemistry and molecular pathology techniques. Although they have high diagnostic value, they are also subjective and limited. With the development of digital pathological images and artificial intelligence technology, traditional typing methods are gradually merging with new technologies. To this end, this application aims to provide an IONet model for ovarian cancer pathological typing based on digital pathological images to address the above problems, improve the accuracy, efficiency and consistency of diagnosis, and provide patients with more precise personalized treatment. Summary of the Invention

[0005] The purpose of the present invention is to solve the technical problems raised in the above background technology and to provide an IONet model for ovarian cancer pathological typing based on digital pathological images.

[0006] The above-mentioned purpose of the present invention is achieved like this:

[0007] On one hand, the solution of the present invention provides a method for pathological typing of ovarian cancer based on digital pathological images. The method is based on pathological images and uses a designed IONet model to classify ovarian cancer, comprising the following steps:

[0008] S1. Automatic background filtering:

[0009] The IONet model was used to perform automated background filtering during WSI processing;

[0010] S2. Multi-scale feature extraction:

[0011] The IONet model is used to optimize computation through downsampling and a multi-scale strategy based on DINO-ViT is used for feature extraction to capture details at the cellular and tissue levels at different resolutions.

[0012] S3, Transformer encoder and MIL classification:

[0013] The IONet model was used for feature optimization using the Transformer encoder and combined with the ABMIL and DSMIL models for classification of ovarian cancer subtypes;

[0014] S4. Integration strategy based on the inference stage:

[0015] The IONet model adopts a simple integration strategy to output the final result by taking a weighted average of the prediction results of each model in step S3.

[0016] Furthermore, the weight of each model in step S4 is 0.25, wherein the ABMIL patch sizes are 8 and 16; and the DSMIL patch sizes are 8 and 16.

[0017] Furthermore, the specific method of automatic background filtering in step S1 is:

[0018] 1) Extract 448x448 image patches from WSI and downsample them by a factor of 0.5;

[0019] 2) The IONet model is used to calculate the ratio of black and white pixels in the patch to evaluate whether each patch belongs to the background area;

[0020] 3) If more than 70% of the area in the patch is background, it is considered to contain no valuable information and is directly discarded;

[0021] Formally, background detection can be described as follows:

[0022]

[0023] The background can therefore be represented as a Boolean matrix that is used to calculate the proportion of useful information in an image patch; the percentage of useful information in a patch is calculated as follows:

[0024]

[0025] Among them, the IONet model only retains image patches whose useful information ratio exceeds a predefined threshold.

[0026] Furthermore, the multi-scale feature extraction in step S2 specifically includes the following steps:

[0027] 1) The IONet model uses the DINO-ViT26 model for feature extraction. DINO-ViT is pre-trained through self-supervised learning, enabling it to learn effective visual features without relying on manual annotation. In the IONet model, DINO-ViT is pre-trained on the TCGA dataset, enabling the model to extract robust and widely applicable features.

[0028] 2) Tissue structures in WSI exhibit unique characteristics at different magnifications. To capture multi-scale features, the IONet model uses a multi-scale patch processing strategy. In the IONet model, patches are processed at two levels: the cellular level and the tissue level. This multi-scale approach ensures that the model can extract valuable information at different resolutions, avoiding the loss of important details.

[0029] Among them, the processed patches can be described as:

[0030] X∈R H×W×C ,

[0031] Where H, W, and C represent the height, width, and number of channels of each patch, respectively; in the ViT model, the patch will be reshaped as:

[0032]

[0033] Among them, H, W and C represent the height, width and number of channels of each patch respectively; for a patch size of 8x8, N = 224×224 / 8×8, resulting in 784 small patches; the small patches are then flattened into 1D vectors and passed through a linear projection layer to map them to a D-dimensional vector space, converting the spatial information of each patch into a format suitable for ViT processing, enabling the model to learn and capture the relationship between patches; the resulting D-dimensional vector is combined with the position embedding to preserve the spatial context, and is then input into the Transformer layer.

[0034] Furthermore, the specific steps of step S3 are:

[0035] 1) Once the patches are flattened and projected, they are passed into a Transformer encoder. The Transformer encoder consists of multiple layers, each of which contains two key components: multi-head self-attention and a feedforward neural network. The multi-head self-attention mechanism enables the model to attend to all patches, thereby understanding the relationship between patches and focusing on different parts of the image simultaneously. After the self-attention of the feedforward neural network, each patch vector is processed through a fully connected neural network to further refine and extract information.

[0036] The multi-head self-attention and feed-forward neural network components are followed by layer normalization and residual connections to ensure stable learning;

[0037] 2) The Transformer encoder outputs a set of refined N×384 patch representations, which are passed as input to the MIL model.

[0038] Furthermore, in step S4, during the inference process of the ensemble strategy, a vector-based importance filter is applied to select the most informative patches; the filter prioritizes patches that contribute most to the final prediction, effectively reducing noise and redundancy in the data.

[0039] On the other hand, the solution of the present invention further provides an IONet model for ovarian cancer pathological typing based on digital pathological images, which is used to implement the above-mentioned method for ovarian cancer pathological typing based on digital pathological images.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. In this paper, automated background filtering and multi-scale feature extraction techniques reduce the reliance on manual annotation and facilitate the processing of large-scale WSI. In addition, by leveraging the self-supervised learning capabilities of DINO-ViT, the IONet model can extract meaningful features without relying on manual labeling, simplifying the model development and training process;

[0042] 2. Our method ensures that the model gives equal weight to both cellular and tissue-level views, resulting in a more balanced and comprehensive data analysis. Furthermore, during inference, we apply a vector-based importance filter to select the most informative patches. This filter prioritizes patches that contribute most to the final prediction, effectively reducing noise and redundancy in the data. By focusing on the most relevant patches, this technique not only improves the efficiency and accuracy of the prediction process but also minimizes unnecessary computational overhead.

[0043] 3. The methods and models of the present invention can ensure the accuracy, efficiency and consistency of ovarian cancer pathological classification, facilitating the provision of more precise personalized treatment for patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the method in an embodiment of the present invention (covering three retrospective datasets. Each whole-slide image was preprocessed and segmented into non-overlapping 448×448 pixel tiles and downsampled to 224×224);

[0045] Figure 2 It is the IONET model design and basic model parameter comparison in the embodiment of the present invention;

[0046] Figure 3 is the performance of IONet and the baseline method in the embodiment of the present invention on two external datasets;

[0047] Figure 4 is the performance of ABMIL in the embodiment of the present invention when extended to two independent external data sets;

[0048] Figure 5 is the performance of DSMIL in the embodiment of the present invention when it is extended to two independent external data sets;

[0049] Figure 6 is the performance of TransMIL in the embodiment of the present invention when extended to two independent external datasets;

[0050] Figure 7 is a mapping of the histomorphological phenotypic landscape in the four ovarian cancer subtypes in the embodiment of the present invention;

[0051] Figure 8 is a heat map visualization of the ovarian cancer subtype task in an embodiment of the present invention;

[0052] Figure 9 It is the result of actual proof-of-concept studies in the embodiments of the present invention;

[0053] Figure 10 These are three cases (one case per row) misdiagnosed by IONet due to complex histopathological features in the present invention (from left to right, including: H&E staining image, whole-slice attention heat map, visualization combining the H&E staining image with its corresponding heat map, and complex histopathological features selected in the black box of the corresponding heat map). DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0055] The implementation of the present invention is described in detail below with reference to specific embodiments.

[0056] With reference to the accompanying drawings, there are shown preferred embodiments of the present invention.

[0057] Example: The present invention provides an IONet model and a method for ovarian cancer pathology classification based on digital pathology images. In this example, the IONet model is an automated deep learning model designed specifically for pathology image (WSI) analysis, and is primarily used for ovarian cancer (OC) classification. The IONet-based ovarian cancer pathology classification method includes the following four core steps:

[0058] Step 1: Automatic background filtering:

[0059] The IONet model automates background filtering during WSI processing, eliminating the need for manual labeling and accelerating model processing efficiency.

[0060] Step 2: Multi-scale feature extraction:

[0061] The IONet model is used to optimize calculations through downsampling, and a multi-scale strategy based on DINO-ViT is adopted for feature extraction to capture details at the cellular and tissue levels at different resolutions, thereby improving the WSI processing efficiency and performance.

[0062] Step 3: Transformer encoder and MIL classification:

[0063] The IONet model was used to perform feature optimization using the Transformer encoder, and the ABMIL and DSMIL models were combined to achieve classification of OC subtypes.

[0064] Step 4: Integration strategy for the inference phase:

[0065] During the inference phase, the model uses a simple ensemble strategy to output the final result by taking a weighted average of the predictions of each model. The weight of all models is 0.25 (ABMIL patch sizes are 8 and 16; DSMIL patch sizes are 8 and 16).

[0066] The specific implementation process of the above implementation plan is as follows:

[0067] 1. Image preprocessing and background filtering

[0068] In traditional pathology image (WSI) processing, segmentation is a key step that usually requires pathologists to manually annotate regions of interest (ROI). IONet automates this process through background filtering.

[0069] The workflow is as follows:

[0070] (1) Extract 448×448 image patches from WSI and downsample them by a factor of 0.5.

[0071] (2) The IONet model is used to evaluate whether each patch belongs to the background area by calculating the ratio of black and white pixels in the patch.

[0072] (3) If more than 70% of the area in the patch is background, it is considered to contain no valuable information and will be discarded directly.

[0073] Formally, background detection can be described as follows:

[0074]

[0075] The background can therefore be represented as a Boolean matrix that is used to calculate the proportion of useful information in an image patch. The percentage of useful information in a patch is calculated as follows:

[0076]

[0077] The IONet model only retains image patches whose proportion of useful information exceeds a predefined threshold. This filtering process not only reduces irrelevant data but also significantly speeds up the model's processing speed.

[0078] A key advantage of automated background filtering is that it completely eliminates the need for manual annotation. While traditional WSI segmentation workflows require pathologists to individually inspect and delineate tissue boundaries for each patch, the IONet model automatically performs this task through simple pixel-wise analysis. By eliminating significant background regions, the IONet model effectively reduces dataset size, thereby improving the efficiency of subsequent model processing.

[0079] 2. Patch downsampling and feature extraction

[0080] After background filtering, the remaining patches are downsampled to a resolution of 224×224. This step optimizes computational resources while preserving critical information. Due to the large size of WSI images, directly processing 448×448 patches would incur significant computational overhead. By reducing the patch size to 224×224, IONet strikes a balance between preserving important information and reducing computational complexity, resulting in significantly faster processing.

[0081] Next, the IONet model uses the DINO-ViT26 model for feature extraction. DINO-ViT is a Visual Transformer (ViT)-based model that is pre-trained through self-supervised learning (SSL), enabling it to learn effective visual features without relying on manual annotation. In the IONet model, DINO-ViT is pre-trained on the TCGA dataset, enabling it to extract robust and widely applicable features, improving its performance across various tasks.

[0082] Tissue structures in WSI exhibit unique characteristics at different magnifications. For example, at higher magnifications, cellular-level features are more pronounced, while at lower magnifications, tissue-level structures are more distinct. To capture these multi-scale features, the IONet model implements a multi-scale patch processing strategy. In the IONet model, patches are processed at two levels: cellular level (8x8 patches) and tissue level (16x16 patches). This multi-scale approach ensures that the model can extract valuable information at different resolutions, avoiding the loss of important details.

[0083] Mathematically, the processed patch can be described as:

[0084] X∈R H×W×C ,

[0085] Where H, W, and C represent the height, width, and number of channels of each patch, respectively. In the ViT model, the patch is reshaped as:

[0086]

[0087] Here, H, W, and C represent the height, width, and number of channels of each patch, respectively. For an 8x8 patch size, we set N = 224x224 / 8x8, resulting in 784 small patches. These small patches are then flattened into 1D vectors and passed through a linear projection layer to map them to a D-dimensional vector space. This step converts the spatial information of each patch into a format suitable for ViT processing, allowing the model to learn and capture the relationship between patches. The resulting D-dimensional vector is combined with the position embedding to preserve the spatial context and is then input into the Transformer layer.

[0088] 3. Transformer Encoder

[0089] Once the patches are flattened and projected, they are fed into a Transformer encoder. The Transformer encoder consists of multiple layers, each of which contains two key components: multi-head self-attention and a feed-forward neural network.

[0090] Multi-head self-attention: This mechanism enables the model to attend to all patches, thereby understanding the relationship between patches and focusing on different parts of the image simultaneously.

[0091] Feedforward Neural Network: After self-attention, each patch vector is processed through a fully connected neural network to further refine and extract information.

[0092] Both components are followed by layer normalization and residual connections to ensure stable learning. The Transformer encoder outputs a set of refined N×384 patch representations, which are passed as input to the MIL model.

[0093] 4. Multi-instance learning and model training

[0094] In the context of multiple instance learning (MIL), each whole slide image (WSI) is represented as a bag of instances, denoted as X = {x1, x2, ..., xk}, where k represents the number of patches derived from the WSI. In a binary classification task, the labels ybag∈{0,1}. In our study, for a four-class classification problem distinguishing ovarian cancer subtypes: HGSC, LGSC, ECOC, and CCOC, the labels ybag=[y1, y2, y3, y4]. After generating multi-scale feature maps, these features are fed into the ABMIL and DSMIL models for training and prediction using the IONet model.

[0095] Both ABMIL and DSMIL are MIL-based methods that effectively manage the imbalanced distribution of patches and adapt to the complex organizational structure of WSIs. ABMIL utilizes an attention mechanism to assign weights to each patch, automatically focusing on the areas most critical for classification. This allows the IONet model to dynamically prioritize the most informative patches. This approach improves model performance. Compared to traditional WSI analysis methods, the IONet model not only improves classification performance but also significantly accelerates processing speed. Figure 3 A comparison of the total parameters of the three models is presented, and the results show that ABMIL and DSMIL require significantly less computational resources, allowing them to run on GPUs with limited memory (e.g. Figure 2 shown).

[0096] Automated background filtering and multi-scale feature extraction techniques reduce reliance on manual annotation and facilitate the processing of large-scale WSI. In addition, by leveraging the self-supervised learning capabilities of DINO-ViT, the IONet model can extract meaningful features without relying on manual labels, simplifying the model development and training process.

[0097] 5. Integration Strategy

[0098] During inference, we employed a simple ensemble strategy by averaging the predictions of each model, assigning an equal weight of 0.25 to all models (ABMIL patch sizes 8,16; DSMIL patch sizes 8,16). This approach ensured that models gave equal weight to both cellular and tissue-level views, resulting in a more balanced and comprehensive data analysis. Furthermore, during inference, we applied a vector-based importance filter to select the most informative patches. This filter prioritized patches that contributed most to the final prediction, effectively reducing noise and redundancy in the data. By focusing on the most relevant patches, this technique not only improved the efficiency and accuracy of the prediction process but also minimized unnecessary computational overhead.

[0099] Regarding IONet model training and evaluation in the implementation scheme of the present invention:

[0100] The implementation scheme of the present invention follows the standard deep learning workflow and does not perform hyperparameter tuning or grid search. The implementation scheme of the present invention randomly divides the internal dataset into training and test sets with a ratio of 4:1 (random seed of 42), and trains each MIL model for 20 epochs using different patch sizes and a batch size of 1. The initial learning rate is set to 5x10^-4, the weight decay is 5x10^-5, and the Adam optimizer is used for optimization. A cosine annealing learning rate scheduler is used to update the learning rate after each epoch to help the model escape the local minimum. We use cross-entropy loss as the loss function and calculate a separate loss for each category. During inference, we use a simple ensemble strategy by averaging the prediction results of each model and giving all models an equal weight of 0.25 (ABMIL patch size 8 and 16; DSMIL patch size 8 and 16). All models are trained on a single RTX4090 GPU.

[0101] Performance evaluation of the model in the embodiment of the present invention:

[0102] The model is evaluated using four metrics: area under the curve (AUC), accuracy (ACC), F1-macro average (F1-macro), and balanced accuracy (BACC). In addition, the embodiment of the present invention also visualizes the total number of parameters of the three models in each view to evaluate their efficiency. In this task, the embodiment of the present invention pays more attention to BACC because the main task of the embodiment of the present invention is to identify the subtypes of ovarian cancer (OC), not just to classify the presence of cancer. Taking into account the natural imbalance of the dataset itself, in which rarer subtypes are less represented, BACC was selected as the key evaluation metric. BACC provides a more comprehensive measure of model performance across all categories, making it particularly suitable in the embodiment of the present invention.

[0103] In this embodiment, Figure 1 In the model, these tiles are further segmented into finer mini-tiles to capture cellular and tissue-level details. A multi-scale feature learning module driven by the Transformer architecture then processes these instance features to produce a comprehensive visual representation of the region of interest (ROI). An instance feature aggregation module treats the ROI as a single instance, applies class-specific attention-weighted aggregation, and generates slice-level predictions. The performance of IONet and comparison methods was initially developed and validated using an internal dataset and subsequently evaluated on two external retrospective datasets. A proof-of-concept dataset was used to evaluate the effectiveness of IONet in ovarian cancer subtype classification. Finally, to enhance the interpretability of the model, IONet was examined for mapping tissue morphological phenotypes between four ovarian cancer subtypes, revealing histological structural features unique to each subtype.

[0104] Figure 2 That is to say Figure 1 The supplement is an overview of the IONet model design and a comparison of basic model parameters, which provides an overview of the entire IONet method. IONet's method starts with extracting 448×448 patches, selecting those containing relevant information through regions of interest (ROIs), and then downsampling to 224×224. These ROIs are then processed by the DINO-ViT model, which is pre-trained and fine-tuned on a large pathology dataset, to further segment the patches into detailed micro-patches at the cellular and tissue scales. The resulting features are processed by a multi-scale feature learning module using a Transformer architecture to generate a robust visual representation of the ROI. Finally, an instance feature aggregation module, which adopts a class-specific attention weighted aggregation paradigm, produces slice-level predictions. Among them, Figure 2 The histogram shown in (b) shows that DSMIL and ABMIL have significantly fewer parameters than TransMIL, reducing the computational cost, especially in terms of GPU usage.

[0105] Figure 3 is the performance of IONet and baseline methods on two external datasets. Figure 3Middle: a. ROC curves and corresponding AUCs for the ovarian cancer subtyping task on the two merged external data. b. Performance indicators of ovarian cancer subtyping by IONet and the baseline method on the two merged external data. c. Confusion matrix of ovarian cancer subtyping by IONet and the baseline method on the two merged external data. d. ROC curves and corresponding AUCs for the ovarian cancer subtyping task on external data 1. e. Performance indicators of ovarian cancer subtyping by IONet and the baseline method on external data 1. f. Confusion matrix of ovarian cancer subtyping by IONet and the baseline method on external data 1. g. ROC curves and corresponding AUCs for the ovarian cancer subtyping task on external data 2. h. Performance indicators of ovarian cancer subtyping by IONet and the baseline method on external data 2. i. Confusion matrix of ovarian cancer subtyping by IONet and the baseline method on external data 2.

[0106] Figure 4 For Figure 3 In addition, we provide an overview of the performance of ABMIL when generalized to two independent external datasets. Figure 4 Middle: a. ROC curve and corresponding AUC for the ovarian cancer subtyping task on the two combined external test datasets. b. ABMIL's confusion matrix for ovarian cancer subtyping on the two combined external test datasets. c. ROC curve and corresponding AUC for the ovarian cancer subtyping task on external test dataset 1. d. ABMIL's confusion matrix for ovarian cancer subtyping on external test dataset 1. e. ROC curve and corresponding AUC for the ovarian cancer subtyping task on external test dataset 2. f. ABMIL's confusion matrix for ovarian cancer subtyping on external test dataset 2.

[0107] Figure 5 For Figure 3 In addition, when DSMIL is extended to two independent external datasets, its performance is described as follows: a. ROC curve and corresponding AUC for the ovarian cancer subtyping task on the two combined external test datasets. b. DSMIL's confusion matrix for ovarian cancer subtyping on the two combined external test datasets. c. ROC curve and corresponding AUC for the ovarian cancer subtyping task on external test dataset 1. d. DSMIL's confusion matrix for ovarian cancer subtyping on external test dataset 1. e. ROC curve and corresponding AUC for the ovarian cancer subtyping task on external test dataset 2. f. DSMIL's confusion matrix for ovarian cancer subtyping on external test dataset 2.

[0108] Figure 6 For Figure 8 The supplement is: when TransMIL is extended to two independent external datasets, its performance is Figure 6 Middle: a. ROC curve and corresponding AUC for the ovarian cancer subtyping task on the two combined external test datasets. b. Average normalized confusion matrix for TransMIL’s classification of ovarian cancer subtypes on the two combined external test datasets. c. ROC curve and corresponding AUC for the ovarian cancer subtyping task on external test dataset 1. d. Confusion matrix for TransMIL’s classification of ovarian cancer subtypes on external test dataset 1. e. ROC curve and corresponding AUC for the ovarian cancer subtyping task on external test dataset 2. f. Confusion matrix for TransMIL’s classification of ovarian cancer subtypes on external test dataset 2.

[0109] Figure 3 is a mapping of the histomorphological phenotypic landscape in four ovarian cancer subtypes. Figure 3 Middle: a. Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction of patches of the four ovarian cancer subtypes within the region of interest within each slide, represented by cluster membership (each cluster is assigned a unique color for enhanced visualization). b. A bar chart illustrates the composition of the various clusters within the four different ovarian cancer subtypes. c. Patches within specific clusters (clusters 3, 4, 5, and 9) exhibit typical histological architectural features of the four ovarian cancer subtypes. Figure 8 .Heatmap visualization of the ovarian cancer subtype task. a. The sequence of four images from left to right includes: 1. A representative slice selected from the HGSC category to generate the corresponding visualization results. 2. Generating a full-slice attention heatmap for each slice by calculating the attention score of the HGSC category. 3. Visualization of merging the representative slice image with its corresponding HGSC category heatmap. 4. Patches with high attention scores in the representative slice effectively reflect the typical morphological features of the HGSC category. b. The visualization results of the CCOC subtype of ovarian cancer are similar to the four images from left to right in the first row. c. The visualization results of the LGSC subtype of ovarian cancer are similar to the four images from left to right in the first row. d. The visualization results of the ECOC subtype of ovarian cancer are similar to the four images from left to right in the first row.

[0110] Figure 9Results from a real-world proof-of-concept study: The proof-of-concept cohort was used to evaluate the performance of IONet, individual pathologists, and the IONet-pathologist duo. Three pathologists participated in the evaluation: a junior pathologist with one year of diagnostic experience, a mid-career pathologist with five years of experience, and a senior pathologist with up to ten years of experience. Two evaluation modes were used: pathologist (unassisted) and pathologist (assisted), the latter of which included assistance from IONet. a. Performance was measured using the AUROC (area under the receiver operating characteristic curve). The IONet-pathologist duo was also evaluated. An arrow line connects the paired performance metrics for the junior, mid-career, and senior pathologists. b. The macro-averaged F1 scores for IONet, pathologist (unassisted), and pathologist (assisted) are plotted for the ovarian cancer subtyping task. c. Confusion matrices are plotted for ovarian cancer subtyping by pathologist (unassisted) and pathologist (assisted) on two external test datasets.

[0111] Figure 10 For Figure 9 Supplementary information: Three cases misdiagnosed by IONet due to complex histopathological features (one case per row). The order of the four images, from left to right, includes: H&E-stained images, whole-slide attention heatmaps. Visualization of the merged H&E-stained images and their corresponding heatmaps, as well as the complex histopathological features selected within the black boxes of the corresponding heatmaps. a. One case diagnosed as a mixed carcinoma, confirmed by morphology and immunohistochemistry. b. The two components of the mixed carcinoma are HGSC (top two) and ECOC (bottom two). c. One case diagnosed as HGSC, confirmed by immunohistochemistry. d. Features of HGSC in most areas (top two) and morphological mimicry of ECOC in some areas (bottom two). e. One case was considered an intermediate carcinoma due to lack of immunohistochemical evidence. f. Features of CCOC in most areas (top two) and focally intermediate morphologies between CCOC and ECOC (bottom two).

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for pathological typing of ovarian cancer based on digital pathological images, characterized in that: Based on pathological images, the designed IONet model is used to classify ovarian cancer, including the following steps: S1. Automatic background filtering: The IONet model was used to perform automated background filtering during WSI processing; S2. Multi-scale feature extraction: The IONet model is used to optimize computation through downsampling and a multi-scale strategy based on DINO-ViT is used for feature extraction to capture details at the cellular and tissue levels at different resolutions. S3, Transformer encoder and MIL classification: The IONet model was used for feature optimization using the Transformer encoder and combined with the ABMIL and DSMIL models for classification of ovarian cancer subtypes; S4. Integration strategy based on the inference stage: The IONet model adopts a simple integration strategy to output the final result by taking a weighted average of the prediction results of each model in step S3.

2. The method for pathological typing of ovarian cancer based on digital pathological images according to claim 1, characterized in that: The weight of each model in step S4 is 0.25, where the ABMIL patch size is 8 and 16; the DSMIL patch size is 8 and 16.

3. The method for pathological typing of ovarian cancer based on digital pathological images according to claim 1, characterized in that: The specific method of automatic background filtering in step S1 is: 1) Extract 448x448 image patches from WSI and downsample them by a factor of 0.5; 2) The IONet model is used to calculate the ratio of black and white pixels in the patch to evaluate whether each patch belongs to the background area; 3) If more than 70% of the area in the patch is background, it is considered to contain no valuable information and is directly discarded; Formally, background detection can be described as follows: The background can therefore be represented as a Boolean matrix that is used to calculate the proportion of useful information in an image patch; the percentage of useful information in a patch is calculated as follows: Among them, the IONet model only retains image patches whose useful information ratio exceeds a predefined threshold.

4. The method for pathological typing of ovarian cancer based on digital pathological images according to claim 1, characterized in that: The multi-scale feature extraction in step S2 specifically includes the following steps: 1) The IONet model uses the DINO-ViT26 model for feature extraction. DINO-ViT is pre-trained through self-supervised learning, enabling it to learn effective visual features without relying on manual annotation. In the IONet model, DINO-ViT is pre-trained on the TCGA dataset, enabling the model to extract robust and widely applicable features. 2) Tissue structures in WSI exhibit unique characteristics at different magnifications. To capture multi-scale features, the IONet model uses a multi-scale patch processing strategy. In the IONet model, patches are processed at two levels: the cellular level and the tissue level. This multi-scale approach ensures that the model can extract valuable information at different resolutions, avoiding the loss of important details. Among them, the processed patches can be described as: X∈R H×W×C , Where H, W, and C represent the height, width, and number of channels of each patch, respectively; in the ViT model, the patch will be reshaped as: Among them, H, W and C represent the height, width and number of channels of each patch respectively; for a patch size of 8x8, N = 224×224 / 8×8, resulting in 784 small patches; the small patches are then flattened into 1D vectors and passed through a linear projection layer to map them to a D-dimensional vector space, converting the spatial information of each patch into a format suitable for ViT processing, enabling the model to learn and capture the relationship between patches; the resulting D-dimensional vector is combined with the position embedding to preserve the spatial context, and is then input into the Transformer layer.

5. The method for pathological typing of ovarian cancer based on digital pathological images according to claim 4, characterized in that: The specific steps of step S3 are: 1) Once the patches are flattened and projected, they are passed into a Transformer encoder. The Transformer encoder consists of multiple layers, each of which contains two key components: multi-head self-attention and a feedforward neural network. The multi-head self-attention mechanism enables the model to attend to all patches, thereby understanding the relationship between patches and focusing on different parts of the image simultaneously. After the self-attention of the feedforward neural network, each patch vector is processed through a fully connected neural network to further refine and extract information. The multi-head self-attention and feed-forward neural network components are followed by layer normalization and residual connections to ensure stable learning; 2) The Transformer encoder outputs a set of refined N×384 patch representations, which are passed as input to the MIL model.

6. The method for pathological typing of ovarian cancer based on digital pathological images according to claim 1, characterized in that: In step S4, during the inference process of the ensemble strategy, a vector-based importance filter is applied to select the most informative patches; the filter prioritizes the patches that contribute most to the final prediction, effectively reducing noise and redundancy in the data.

7. An IONet model for ovarian cancer pathological typing based on digital pathological images, characterized in that: A method for pathological typing of ovarian cancer based on digital pathological images is used to implement any one of claims 1-6.

Images