Reorganized MIDOG++: A Patch-level Dataset for Mitotic Figure Detection and Segmentation

Dataset Summary This repository provides a reorganized, patch-level version of the MIDOG++ (Mitosis Domain Generalization) dataset. It is formatted for object detection (YOLO) and instance segmentation. The data is derived from all 503 high-resolution Regions of Interest (ROIs) included in the original collection, spanning diverse cancer types across human and canine tissue. All ROIs were tiled into standardized 512x512 pixel patches. Data Processing & Features Exhaustive Tiling: All 503 ROIs were processed without discarding tissue areas. Zero-Loss Padding: Black padding was applied to edge patches where ROI dimensions were not perfectly divisible by 512, preserving 100% of the original spatial data. Annotated Labels: Each patch includes a corresponding YOLO-format annotation file. CellSAM Masks: Mitotic figure masks were generated using a CellSAM pipeline and preprocessed to provide fine-grained morphological boundaries. Repository Structure MIDOGpp_Patches/: Image patches (PNG). MIDOGpp_Labels/: YOLO-format .txt annotation files. MIDOGpp_Masks/: Binary .png masks (0: Background, 255: Mitosis). metadata.csv: Master file mapping patches to Patient IDs and mask availability.