Non-relevant segment recognition via hard example mining under sparsely distributed events

Area of Science:

• Medical Image Analysis
• Surgical Technology
• Machine Learning in Healthcare

Background:

• Non-relevant segment (NRS) recognition in surgery is crucial for safety and efficiency.
• Sparse event distribution in surgical data challenges the generalization performance of recognition models.
• Existing performance metrics may not adequately address the critical nature of false positives in surgical contexts.

Purpose of the Study:

• To introduce on/offline hard example mining (HEM) techniques to improve NRS recognition performance.
• To develop and validate novel performance metrics suitable for clinical NRS recognition tasks.
• To demonstrate the utility of the proposed methods in long-duration surgeries and diverse surgical environments.

Main Methods:

• Implemented on/offline hard example mining (HEM) to extract challenging NRS training examples.
• Introduced False Discovery Rate (FDR) and Threat Score (TS) as quantitative performance metrics.
• Validated the methodology on long-duration robotic and laparoscopic surgery datasets.

Main Results:

• HEM significantly improved NRS recognition performance by effectively training on difficult examples.
• FDR and TS provided clinically relevant quantitative evaluations, addressing limitations of traditional metrics.
• The proposed model demonstrated robust utility and applicability in long-duration surgeries and various clinical settings.

Conclusions:

• On/offline HEM is an effective strategy to enhance NRS recognition, particularly in sparse data scenarios.
• FDR and TS are valuable metrics for evaluating NRS recognition in clinical practice, especially where false positives are critical.
• The developed methodology shows promise for improving surgical safety and efficiency across different surgical durations and modalities.