Patch-Level Glioblastoma Subregion Classification with a Contrastive Learning-Based Encoder

TL;DR

The paper addresses patch level subregion classification in glioblastoma histopathology under strong heterogeneity. It employs a Virchow2 Vision Transformer backbone pretrained on histopathology via self supervised learning (DINOv2) and a dual representation that combines global meanToken and classToken into a $2560$-dimensional feature, feeding a $256$-d bottleneck to predict $9$ classes. On BraTS-Pathology 2025, the method achieves online validation MCC $0.7064$ and F1 $0.7676$, and a final test MCC $0.6509$ with F1 $0.5330$, securing second place. These results establish a solid baseline for ViT based histopathology and highlight generalization gaps for unseen data, pointing to future work in data augmentation, regularization, and long-tailed class handling.

Abstract

The significant molecular and pathological heterogeneity of glioblastoma, an aggressive brain tumor, complicates diagnosis and patient stratification. While traditional histopathological assessment remains the standard, deep learning offers a promising path toward objective and automated analysis of whole slide images. For the BraTS-Path 2025 Challenge, we developed a method that fine-tunes a pre-trained Vision Transformer (ViT) encoder with a dedicated classification head on the official training dataset. Our model's performance on the online validation set, evaluated via the Synapse platform, yielded a Matthews Correlation Coefficient (MCC) of 0.7064 and an F1-score of 0.7676. On the final test set, the model achieved an MCC of 0.6509 and an F1-score of 0.5330, which secured our team second place in the BraTS-Pathology 2025 Challenge. Our results establish a solid baseline for ViT-based histopathological analysis, and future efforts will focus on bridging the performance gap observed on the unseen validation data.

Figures (5)

• Figure 1: Virchow2 uses four registers to mitigate local information loss and enhance global contextual information by storing artifact tokens, thereby enabling better feature extraction. Adapted from darcet2024.
• Figure 2: The network architecture of our model.
• Figure 3: Illustration of the annotated histologic areas of interest.
• Figure 4: Class distribution of the dataset, detailing the number and percentage of samples for each category.
• Figure 5: Aggregated confusion matrix from the 5-fold cross-validation on the local validation set.