Topology-guided hard example mining for cell detection.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Imaging Analysis

Background:

• Manual cell counting in digital pathology is time-consuming and error-prone.
• Current deep learning models struggle with accurate cell detection in crowded images.
• Existing methods often overlook the critical global organization and topological structure of cell arrangements.

Purpose of the Study:

• To introduce a novel training strategy, topology-guided hard example mining (TG-HEM), for improved cell detection.
• To address the limitations of existing methods in handling crowded cell distributions by incorporating topological constraints.
• To enhance the capture of higher-order organization in cell distributions within digital pathology images.

Main Methods:

• TG-HEM is a training strategy using loss reweighting to incorporate topological constraints.
• It quantifies topological discrepancies using persistent homology, identifying challenging regions beyond pixel-wise errors.
• The approach reweights regions with topological inconsistencies and hard-to-learn pixels, guiding network training.

Main Results:

• TG-HEM consistently improved cell counting and localization accuracy across multiple network architectures.
• Performance gains were observed on both the BRCA-M2C and the in-house KUCell datasets.
• The method demonstrated improvements without increasing model complexity, inference time, or significantly impacting training time.

Conclusions:

• TG-HEM offers an effective approach to enhance deep learning-based cell detection in digital pathology.
• The strategy successfully addresses challenges posed by crowded cell distributions by leveraging topological information.
• This method provides a valuable tool for more accurate and reliable automated cell analysis in pathological images.