PseudoCell: Hard Negative Mining as Pseudo Labeling for Deep Learning-Based Centroblast Cell Detection.

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Deep learning models aid follicular lymphoma grading using whole-slide images (WSIs).
• Current models necessitate manual centroblast identification and refined labels by pathologists for optimization.
• This requirement poses a significant bottleneck in the diagnostic workflow.

Purpose of the Study:

• Introduce PseudoCell, an object detection framework for automated centroblast detection in WSIs.
• Eliminate the need for extensive pathologist-refined labels, streamlining model optimization.
• Reduce the manual annotation burden in digital pathology.

Main Methods:

• Employ a hybrid approach combining pathologist-provided centroblast labels with pseudo-negative labels.
• Generate pseudo-negative labels from undersampled false-positive predictions utilizing cell morphology.
• Develop an object detection framework for precise centroblast identification.

Main Results:

• PseudoCell significantly reduces pathologist workload by accurately identifying areas of interest.
• The framework can eliminate 58.18-99.35% of irrelevant tissue areas in WSIs, depending on the confidence threshold.
• Demonstrate PseudoCell's efficiency in prescreening for centroblast detection.

Conclusions:

• PseudoCell offers a practical and efficient prescreening solution for centroblast detection in WSIs.
• The framework obviates the need for refined pathologist labels, enhancing workflow efficiency.
• Provide guidance for clinical implementation of PseudoCell in diagnostic settings.