MaskMitosis: a deep learning framework for fully supervised, weakly supervised, and unsupervised mitosis detection in...

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning for cancer diagnosis

Background:

• Accurate counting of mitotic cells in histopathological images is vital for aggressive breast cancer grading.
• Existing methods face challenges with precise mitosis detection and segmentation in diverse datasets.

Purpose of the Study:

• To develop a robust and accurate automatic mitosis detection and instance segmentation technique for histological breast cancer slides.
• To leverage the Mask R-CNN framework for both fully and weakly annotated datasets.

Main Methods:

• Utilized the Mask R-CNN multi-task deep learning framework for object detection and instance segmentation.
• Employed a two-stage approach for weakly annotated datasets, first generating pseudo-labels from a pre-trained model.
• Evaluated the framework on the 2012 ICPR (fully annotated) and 2014 ICPR MITOS-ATYPIA (weakly annotated) challenge datasets.

Main Results:

• Achieved an F-score of 0.863 on the fully annotated 2012 ICPR dataset.
• Outperformed state-of-the-art methods on the weakly annotated 2014 ICPR dataset with an F-score of 0.475, using a two-stage training strategy.
• Demonstrated promising unsupervised mitosis detection capabilities through pseudo-label estimation.

Conclusions:

• The proposed Mask R-CNN framework offers a robust and accurate solution for automatic mitosis detection and segmentation in breast cancer histology.
• The two-stage approach effectively handles weakly annotated data, improving performance on challenging datasets.
• The method shows potential for unsupervised analysis and contributes to advancing computational pathology tools.