Weakly supervised mitosis detection in breast histopathology images using concentric loss

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computational Pathology

Background:

• Accurate mitosis counting is crucial for breast cancer grading and diagnosis.
• Manual mitosis counting by pathologists is labor-intensive and time-consuming.
• Existing deep learning methods struggle with weakly labeled histopathology data.

Purpose of the Study:

• To develop an automated deep learning system for mitosis detection in breast cancer histopathology images.
• To address the challenge of training segmentation models with weak centroid labels.
• To improve the efficiency and accuracy of breast cancer grading.

Main Methods:

• Proposed a deep learning scheme using a fully convolutional network for semantic segmentation.
• Introduced a novel "concentric loss" function to handle weak centroid pixel labels.
• Expanded single-pixel labels to concentric circles, excluding the uncertain 'middle ground' ring from loss calculation.

Main Results:

• Achieved state-of-the-art F-scores on multiple benchmark datasets: 0.562 (ICPR 2014 MITOSIS), 0.673 (AMIDA13), and 0.669 (TUPAC16).
• Demonstrated significant performance improvement over previous mitosis detection approaches.
• Validated the effectiveness of the proposed concentric loss for weakly supervised learning.

Conclusions:

• The proposed deep learning method with concentric loss enables effective mitosis segmentation using weakly annotated data.
• Automated mitosis detection can significantly aid pathologists in breast cancer diagnosis and grading.
• The publicly available code facilitates further research and development in the field.