Classifying Mitotic Figures in the MIDOG25 Challenge with Deep Ensemble Learning and Rule Based Refinement
This work addresses the challenge of automating mitotic figure classification for tumor grading, which is incremental as it builds on existing deep learning methods with minor modifications.
The paper tackled the problem of classifying mitotic figures in tumor grading by developing an ensemble of ConvNeXtBase models with rule-based refinement, achieving a balanced accuracy of 84.02% on the MIDOG25 test set, though the refinement reduced overall performance.
Mitotic figures (MFs) are relevant biomarkers in tumor grading. Differentiating atypical MFs (AMFs) from normal MFs (NMFs) remains difficult, as manual annotation is time-consuming and subjective. In this work an ensemble of ConvNeXtBase models was trained with AUCMEDI and extend with a rule-based refinement (RBR) module. On the MIDOG25 preliminary test set, the ensemble achieved a balanced accuracy of 84.02%. While the RBR increased specificity, it reduced sensitivity and overall performance. The results show that deep ensembles perform well for AMF classification. RBR can increase specific metrics but requires further research.