ReCasNet: Improving consistency within the two-stage mitosis detection framework.

Area of Science:

• Digital pathology
• Computational biology
• Machine learning in healthcare

Background:

• Manual mitotic count (MC) is crucial for cancer diagnosis but is time-consuming and error-prone.
• Existing deep learning models for MC analysis use a two-stage detection-classification pipeline with inherent inconsistencies.
• These inconsistencies arise from poor detection quality and mismatched training data distributions between stages.

Purpose of the Study:

• To introduce Refine Cascade Network (ReCasNet), an improved deep learning pipeline for accurate mitotic cell detection and MC prediction.
• To address limitations of existing two-stage pipelines, including false positives and data distribution mismatches.
• To enhance the efficiency and reliability of automated mitotic cell analysis in histopathology.

Main Methods:

• Developed ReCasNet, a novel deep learning pipeline incorporating window relocation and object re-cropping.
• Implemented improved data selection strategies for the classification stage to mitigate training data distribution mismatches.
• Evaluated ReCasNet on canine cutaneous mast cell tumor (CCMCT) and canine mammary carcinoma (CMC) datasets.

Main Results:

• ReCasNet achieved up to 4.8% percentage point improvements in F1 scores for mitotic cell detection.
• Demonstrated up to 44.1% reductions in mean absolute percentage error (MAPE) for MC prediction.
• Showcased the effectiveness of proposed techniques in improving performance on large-scale datasets.

Conclusions:

• ReCasNet significantly enhances mitotic cell detection and MC prediction accuracy compared to existing methods.
• The ReCasNet pipeline offers a robust solution for automating MC analysis in digital pathology.
• The underlying techniques are generalizable to other two-stage object detection pipelines, advancing AI in pathology.