Sequential Hard Mining: a data-centric approach for Mitosis Detection
This work addresses the problem of optimizing deep learning training for mitosis detection in medical imaging, but it appears incremental as it builds on existing methods.
The paper tackles the challenge of efficiently using large annotated datasets for mitosis detection in histology images by proposing a data-centric approach based on sequential hard mining and boosting techniques, with results applied to the MIDOG 2025 challenge tracks.
With a continuously growing availability of annotated datasets of mitotic figures in histology images, finding the best way to optimally use with this unprecedented amount of data to optimally train deep learning models has become a new challenge. Here, we build upon previously proposed approaches with a focus on efficient sampling of training data inspired by boosting techniques and present our candidate solutions for the two tracks of the MIDOG 2025 challenge.