Cell Image Classification with Polygon-Guided Extraction for Glioma Mitosis Detection

Accurate detection of mitotic figures in glioma tissue samples is crucial for assessing tumor proliferation rates and planning appropriate treatment strategies. This paper presents our contribution to the Glioma-MDC 2025 Challenge, introducing an efficient approach that combines precise cell extraction with state-of-the-art deep learning models. Our methodology employs a systematic cell extraction pipeline that maintains cellular detail while standardizing input for classification, coupled with ConvNeXt V2 architecture implementations. The pipeline processes H&E stained image patches and their corresponding polygon annotations, incorporating a 5-pixel margin to ensure complete cellular context capture. We evaluated both ConvNeXt V2 Tiny and Base models across various input dimensions, identifying optimal performance in the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$128\times 224$</tex> pixel range. Our best-performing models achieved a public score of 0.98136 and private score of 0.98550, securing 2nd place among 29 competing teams and demonstrating the effectiveness of our approach. Analysis of the extracted cell regions revealed consistent dimensional patterns, with aspect ratios close to 1.0 across all categories, indicating successful standardization while preserving morphological characteristics. Notable findings include comparable performance between lightweight and full-scale model variants, suggesting potential for widespread clinical deployment. The study provides valuable insights for advancing automated mitotic figure detection in digital pathology, though further validation across diverse institutional settings is recommended for robust clinical implementation.