Reproduce “Patch-Level Glioblastoma Subregion Classification with a Contrastive Learning-Based Encoder” on OpenResearch

About this paper

Juexin Zhang, Qifeng Zhong, Ying Weng, Ke Chen

Researchers from the University of Nottingham Ningbo China developed a deep learning model for patch-level classification of nine glioblastoma subregions, fine-tuning a pre-trained Vision Transformer. The model achieved an F1-score of 0.8775 on local validation and secured second place in the BraTS-Pathology 2025 Challenge, though a notable performance drop was observed on external test sets.

Key insights

• Large-scale pre-trained Vision Transformers are effective in extracting rich, generalizable histopathological features for complex medical image classification tasks.
• Combining global average and semantic class token representations provides a robust feature vector for patch-level classification.
• A notable 'generalization gap' exists between local validation and external challenge test sets, indicating challenges with overfitting and robustness to unseen data distributions in medical AI models.
• Model performance is heavily impacted by class imbalance, with minority classes showing significantly lower recall and F1-scores.

Summary by alphaXiv.