Domain Adaptive Box-Supervised Instance Segmentation Network for Mitosis Detection.

Area of Science:

• Digital pathology
• Computational oncology
• Medical image analysis

Background:

• Accurate mitosis counting in histopathology is crucial for breast cancer diagnosis and prognosis.
• Current methods struggle with pixel-level prediction on weakly labeled mitosis datasets and domain variations across labs.

Purpose of the Study:

• To develop a robust method for precise pixel-level mitosis segmentation using only centroid locations (weak labels).
• To address the domain shift problem in histopathological images from different sources.
• To improve the generalization ability of mitosis detection models.

Main Methods:

• Proposed a Domain adaptive Box-supervised Instance segmentation Network (DBIN).
• Introduced a Box-supervised Instance-Aware (BIA) head with novel mask loss terms.
• Incorporated a Pseudo-Mask-supervised Semantic (PMS) head to enhance feature representation.
• Implemented a Cross-Domain Adaptive Module (CDAM) to align feature distributions between domains.

Main Results:

• Achieved state-of-the-art performance on four mainstream datasets.
• Demonstrated effective pixel-wise mitosis localization under weak supervision using BIA and PMS heads.
• Showcased improved model generalization through the CDAM, enabling adaptation to new data domains.

Conclusions:

• DBIN provides a powerful framework for accurate mitosis segmentation in breast cancer histopathology.
• The proposed BIA and PMS heads enable precise localization with weak supervision.
• CDAM significantly enhances model adaptability and performance across diverse datasets.