Differential cell counts using center-point networks achieves human-level accuracy and efficiency over segmentation.

Area of Science:

• Computational Pathology
• Computer Vision
• Medical Diagnostics

Background:

• Accurate differential cell counts are crucial for pathological diagnosis but challenging for computer vision algorithms.
• Existing methods require extensive annotations (segmentation/bounding boxes) and struggle with clustered cells.

Purpose of the Study:

• To introduce DCNet (Differential Count Network), an annotation-efficient modality for differential cell counting.
• To reduce the expert burden for ground truth data generation in cell recognition tasks.

Main Methods:

• DCNet utilizes keypoint detection to identify cell center points and their classes in brightfield images.
• This method requires only single center point annotations, reducing expert labeling time by 77.1% compared to bounding boxes.

Main Results:

• DCNet achieved high accuracy in multi-class cell recognition using whole cell features.
• Performance matched human experts in average precision across 5 object classes and exceeded human consistency.

Conclusions:

• DCNet offers an efficient and accurate computational approach to differential cell counting.
• This system represents significant progress towards an open-source, automated solution for pathology diagnostics.