At present, the existence of an open database is very important. The complexity of cancer disease requires a large amount of data as a supporting basis, and the existence of an open database can accelerate related research and save researchers’ resources (16). Therefore, this retrospective study included histopathological slides of H&E-stained tissue samples from two cohorts of glioblastoma patients. The first cohort comprised patients from The Cancer Genome Atlas (TCGA) project, accessible through the National Institutes of Health Genome Data Sharing Portal (17–19) (TCGA cohort). The second cohort comprised patients with glioblastoma diagnosed in Beijing Tiantan Hospital from 2019 to 2023 (TianTan cohort). As cryopreservation destroys the original morphological structure of cells and tissues, all histopathological slides collected in this study were preserved by formalin-fixed paraffin embedding. Specific data inclusion and exclusion criteria are listed below:

Inclusion criteria were as follows: (1) precise diagnosis of glioblastoma by surgical pathology; (2) clear methylation status of the MGMT promoter by genetic testing; and (3) no other brain lesions. Exclusion criteria were as follows: (1) unclear images of histopathological slides; and (2) missing baseline information or unknown medical history.

All pathological images in the TCGA cohort were confirmed to be glioblastoma by secondary diagnosis by a pathologist with extensive experience in pathological diagnosis. All histopathological slides in the TianTan cohort were scanned using a Lecia Aperio CS2 scanner into WSIs of digital pathology in standard file format (the “svs” format). The scanning magnification was 20X. All data from the TianTan cohort were only used as an external independent test set to verify the generalization of the model and were not used in the training or hyperparameter adjustment process. Data in the TCGA cohort were divided into the training set and the test set according to certain standards: If multiple WSIs originated from the same patient, then all those WSIs from that patient were assigned to the training set. This approach ensured that slides from the same tumor sample were not used for both training and testing at the same time.

This study was approved by the Ethics Committee of Beijing Tiantan Hospital (ethical approval no: YW2022-025).

Our study achieved end-to-end prediction of MGMT promoter methylation status in glioma patients based on H&E-stained WSIs through three main modules. First, the scanned WSIs were divided into several patches using the preprocessing module. After that, the glioblastoma region of interest (ROI) patches prediction module was used to filter the sliced patches. Finally, the molecular status of patients was predicted by the MGMT promoter status prediction module. The main body model framework adopts the modular design of openMMLab (20). The details are described below. The complete process diagram is shown in Figure 1.

This module is crucial for pre-processing WSIs. It consisted of two parts. The first part is the patch segmentation. This part is responsible for splitting the complete WSI into patches. A complete WSI consists of billions of pixels through a pyramid structure. It is difficult to directly input such large image data into a neural network or machine learning model for training. Therefore, we used the sliding window method to divide the tissue image in WSI into several patches of 512 pixels × 512 pixels for subsequent data analysis and processing (21). The second part is the color normalization. This part is dedicated to the color normalization of the patches. Color variation can arise due to several factors, including variations introduced by the operator during the H&E staining and differences in staining reagents. Such color variations can significantly interfere with subsequent modeling predictions. Thus, we used the Vahadane method to normalize the colors presented in H&E pathology images to eliminate the influence of staining differences, ensuring consistent and reliable data for further analysis (22).

This module is used to screen the patches obtained in the Preprocessing module. As a WSI may encompass diverse elements such as hemorrhage and normal tissue, potentially interfering with the final prediction, we employed the ROI patches prediction model to screen the patches generated in the preprocessing module to avoid the interference of confusing information as much as possible. The module utilized transfer learning, employing a Vision Transformer model pre-trained on the ImageNet dataset (23). We optimized the model’s performance using the AdamW optimizer (24). To further enhance model convergence during training, we employed the CosineAnnealingLrUpdater strategy for optimal learning rate adjustments (25). The initial learning rate is set at 1.75 × 10⁻³, with a minimum of 1.75 × 10⁻⁵ imposed during the learning rate adjustment process. We also employed linear preheating for the learning rate. Model training and prediction used the LabelSmoothLoss as the loss function (26).

In this process, the training data for the Vision Transformer model were derived from the annotations of pathologists. Additionally, pathologists screened patches to ensure that the training dataset exclusively includes ROI patches exhibiting typical tumor characteristics.

In addition, to investigate the potential impact of refined annotation on the results, we explored other two methods to filter the patches generated by the Preprocessing Module. This included relying on the pathologist’s annotation for filtering (Tumor Only) and only removing the blank background (Whole Tissue). The implementation details for each method are described below:

Through the above method, we obtained three groups of patches annotated in different ways. These three groups of patches were input into the final MGMT Predict Module to obtain diverse results and to compare the impact of these three annotation methods on the results.

The ROI patches output by the second module were used as the input of this module to predict the final MGMT promoter methylation status. The processing flow of this module involved considering each WSI as a package, which is then sliced into numerous ROI patches. Each ROI patch is treated as an instance, with the package’s label being assigned to each patch within it. In this part, the main structure of our model adopts the Swin Transformer v2 structure (27). After position encoding of input data using PatchEmbed, the four stages are passed. Except for the first stage, each stage consists of a PatchMerging module and several Swinblock V2 modules. Then, the features are processed by the linear normalization layer and global average pooling layer and then predicted and classified by a fully connected layer. Each SwinBlockV2 module consists mainly of multi-head attention mechanisms with feedforward neural networks and linear normalization layers. In the four stages, the number of SwinBlockV2 is [2, 2, 18, 2], respectively, and the number of attention heads for each SwinBlockV2 is [3, 6, 12, 24], respectively. In this way, for each WSI, we could get the prediction probability of several patches and perform a weighted average of all the prediction results to obtain the final patient-level prediction. This ensured that each patch’s predictions contributed appropriately to the patient-level result. The weighting formula was as follows:

Weigh t i represents the weight that the prediction result of the i-th patch contributes to the final weighted average, and pred _ scor e i signifies the prediction result of the i patch.

For patch-level results, weights were assigned based on the degree of deviation of their prediction probability from 0.5. This strategic weighting reduced the influence of results with prediction probabilities near 0.5 on the final average, thereby strengthening the reliability of model predictions.

We used SPSS 26 to analyze the differences in clinical baseline data of patients using a t-test and chi-square test. Python (3.7.9) was used for data processing and analysis and overall model construction. We employed the PyTorch framework to construct the neural network, and the hardware platform for model training was the NVIDIA A40 GPU with 48G of memory.

According to the inclusion and exclusion criteria, we included 29 WSIs from 29 patients in the TianTan cohort and 632 WSIs from 199 patients in the TCGA cohort. A total of 57 patients in the TCGA database had only one WSI based on the previously described criteria, so 57 WSIs from these 57 patients were used as an independent test set for the TCGA cohort. The remaining 575 WSIs from 142 patients were used as the training set. According to the fifth edition of the WHO Classification of Tumors of the Central Nervous System (WHO CNS5) (5), we excluded patients with IDH mutation from the test set. Therefore, 25 patients with corresponding 25 WSIs were finally included in the TianTan cohort.

There were no statistically significant differences between the two cohorts regarding age, sex, and MGMT promoter methylation status. Basic clinical characteristics are presented in Table 1.

Our primary evaluation metrics included accuracy and the area under the curve (AUC) value of the receiver operating characteristic (ROC) curve. These metrics were used to assess the performance of the models in both the glioblastoma ROI patches prediction module and the MGMT promoter status prediction module.

In the ROI patches recognition module, the training dataset comes from the 100 WSIs in the training set of the TCGA cohort described earlier. After segmentation and color normalization, 14,000 patches with typical characteristics were selected under the guidance of professional pathologists. There were 7,000 patches each in the ROI area and non-ROI area. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio. During training, a batch size of 224 and 100 epochs were used for the parameter settings of the model. The model reached the highest accuracy of 97.88% in the validation set at the 80th epoch, and the model parameters at this time were saved for testing in the test set. The final accuracy of the ROI patches prediction module in the test set is 97.03%, and the AUC value is 0.9940 (Figure 2). The results show that the module has a superior ability to identify ROI patches, which can meet the screening needs of the subsequent modules.

In the MGMT promoter status prediction module, we first used the Swin Transformer V2 model for training. The ROI patches output by the ROI patches prediction module was used as the training data of this module. Whether the patient’s MGMT promoter was methylated or not was taken as the label of the patches cut out from that patient. Therefore, we finally obtained 59,945 MGMT promoter methylated patches and 97,431 MGMT promoter unmethylated patches.

To address the issue of category imbalance, we employed random rotation and mirroring to expand both categories to 100,000, with a total of 200,000 patches for model training. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio. We set the batch size to 148 for training parameters and the number of epochs to 500. Remarkably, during the 493rd epoch, the model achieved its highest accuracy of 91.38% in the validation set. We saved the model weights at this time for testing on the test set. In the end, the model exhibited an accuracy of 91.12% in the test set, with an AUC of 0.98 (Figure 3). At the same time, we compared it with the current commonly used Resnet50 model, and the comparison results are shown in Table 2. These results show the ability of the model to accurately distinguish whether the MGMT promoter is methylated at the patch level.

To enhance the model’s interpretability, we utilized a class gradient activation heatmap to visualize the model’s recognition of patches. In this visualization, regions with a redder color indicate a higher probability that those areas are methylated. As depicted in Figure 4A, represents a WSI with MGMT promoter methylation, while Figure 4B represents a WSI without MGMT promoter methylation. Each image’s upper left corner displays the corresponding patch’s gradient class activation heatmap. The results demonstrate that the models correctly identify their respective patches.

Hence, we aimed to extrapolate the patch-level results to the patient level. The patches obtained by three different methods were input into the MGMT promoter status prediction module for testing. According to the ROC curve of previous patch-level evaluation, we calculated that the best cutoff value for judging whether MGMT promoter methylation exists at the patch level was 0.52; that is, when the probability of methylation of a patch MGMT promoter judged by the model was greater than 0.52, we think it has MGMT promoter methylation. We use this cutoff value for patient-level model evaluation. To obtain patient-level results, we aggregated the patch-level outcomes using the probability-weighted average method as previously described. The results are presented in Table 3.

In both TCGA and TianTan test sets, the approach combining the ROI patches prediction module with the MGMT promoter status prediction module demonstrates the highest AUC values. Moreover, the approach combining the ROI patches prediction Module with the MGMT promoter status in the Tiantan cohort and the TCGA cohort T achieved the same highest balance accuracy. The ROC curves for these three methods in the test set are depicted in Figure 5. These comprehensive results demonstrate that the overall process we designed can effectively determine the presence or absence of MGMT promoter methylation in patients through the H&E-stained histopathological slides.