A Semi-Supervised Learning Framework with Cross-Magnification Attention for Glioma Mitosis Classification

Mitotic figure counting plays a critical role in glioma grading and prognostication, yet manual counting remains time-consuming and subject to variability. The Glioma-MDC 2025 Challenge, hosted at ISBI 2025, aims to address this challenge by advancing automated solutions. In response, we present SEMA, a semi-supervised learning framework that incorporates a cross-magnification attention mechanism. SEMA mimics the multi-magnification workflow used by pathologists, employing attention to model interactions between cells and neighboring regions, effectively replicating how pathologists analyze whole slide images. Additionally, SEMA leverages the abundance of unlabeled cells in raw tissue images from competition datasets through a semi-supervised learning approach, progressively enhancing its robustness and performance. Extensive experiments demonstrate that SEMA outperforms other baselines by up to 40% in FI score, with each component contributing significantly to its success. Notably, SEMA achieves a perfect F1 score, securing top performance on the public leaderboard.