MaskMitosis

MaskMitosis detects and segments mitotic figures in histopathology images to support mitotic counting and grading in breast cancer.

Key Features:

• Multi-Task Deep Learning Framework: Integrates mitosis detection with instance segmentation to enable precise localization and classification of mitotic cells.
• Mask R-CNN Architecture: Implements Mask R-CNN to perform simultaneous object detection and instance segmentation for mitotic figures.
• Fully Supervised Detection: Trained on pixel-level annotations such as the 2012 ICPR Grand Challenge dataset and reported an F-score of 0.863 on that dataset.
• Weakly Supervised Two-Stage Framework: For centroid-pixel labels (e.g., 2014 ICPR MITOS-ATYPIA), uses a model trained on fully annotated data to generate mitosis mask and bounding box labels from weak annotations, then trains a second model on the estimated labels, achieving an F-score of 0.475.
• Unsupervised Detection via Pseudo Labels: Can estimate pseudo labels for unlabeled datasets to enable unsupervised detection.
• Outputs: Produces mitosis masks, bounding boxes, instance-level segmentation, localization, and classification of mitotic cells.

Scientific Applications:

• Tumor Grading in Breast Cancer: Automates mitotic cell counting to support assessment of tumor grade in breast cancer histopathology.
• Diagnostic and Therapeutic Decision Support: Provides quantitative mitosis detection results that can inform diagnostic assessments and treatment planning.

Methodology:

Implements Mask R-CNN for simultaneous detection and instance segmentation; supports fully supervised training on pixel-level annotations, a two-stage weakly supervised pipeline where a model trained on fully annotated data generates mitosis masks and bounding boxes from centroid annotations to train a second detector, and estimation of pseudo labels for unsupervised detection.