Position-based anchor optimization for point supervised dense nuclei detection.

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Nuclei detection is crucial for computer-aided cancer diagnosis but requires extensive manual annotation in fully-supervised methods.
• Weakly-supervised learning offers a solution to reduce annotation burden, yet detecting dense, crowded nuclei remains difficult.
• Existing methods struggle with complex nuclei distributions and diverse appearances.

Purpose of the Study:

• To develop a novel point-supervised framework for accurate dense nuclei detection in histopathological images.
• To reduce the reliance on time-consuming and expert-dependent manual annotations.
• To improve nuclei detection performance in crowded and diverse cellular environments.

Main Methods:

• A point-supervised dense nuclei detection framework utilizing position-based anchor optimization and morphology-based pseudo-label supervision.
• Generation of cellular-level pseudo labels (CPL) using a morphology-based mechanism.
• Implementation of Position-based Anchor-quality Estimation (PAE) to refine detections in crowded areas.
• Introduction of an Adaptive Anchor Selector (AAS) for robust anchor selection based on nuclei characteristics.

Main Results:

• The proposed framework demonstrates superior performance compared to state-of-the-art methods on MO and Lizard benchmarks.
• Achieved 95.1% of fully-supervised performance specifically in dense nuclei detection scenarios.
• Validated effectiveness using ResNet50 and PVTv2 backbones.

Conclusions:

• The novel point-supervised approach effectively addresses the challenges of dense nuclei detection with reduced annotation effort.
• The framework shows significant potential for enhancing computer-aided diagnosis in digital pathology.
• The developed methods offer a practical solution for analyzing complex histopathological images.