Automated Detection of Mitotic Figures in Glioma Tissues Using Deep Learning

Deep learning approaches have shown significant promise in advancing automated detection in digital pathology. In this study, we present a comprehensive investigation of methods for accurately detecting mitotic activity in regions of interest extracted from H&E stained glioma tissue images. Our approach involves training a Vision Transformer (ViT) from scratch while also evaluating alternative models, including ResNet, Swin Transformer, DenseNet, and VGG architectures, supplemented by various transfer learning strategies using ImageNet pretrained weights. Central to our work is the Glioma-MDC 2025 challenge, which serves as a benchmark for developing robust algorithms that detect and classify mitotic figures in glioma tissues. Experimental results, assessed via Accuracy, Precision, Recall, and F1 Score through cross-validation, demonstrate that our transformer based models achieve a highest F1 score of 0.982 on training data and an external test score of 0.969 during the challenge's validation phase. These findings suggest that our carefully curated approaches can significantly enhance automated mitosis detection, potentially leading to more accurate glioma grading in clinical settings.