Evaluation Framework for Bruise Detection: Systematic ALS/White-Light Training and Skin-Tone Balancing with Deep...

Area of Science:

• Forensic Science
• Computer Vision
• Medical Imaging

Background:

• Accurate forensic bruise assessment is vital for victims of violence.
• Existing methods face challenges with equitable performance across skin tones and lighting.

Purpose of the Study:

• To present a framework for automated bruise detection integrating alternate-light-source (ALS) and white light imaging.
• To address equitable performance across skin tones and lighting scenarios using novel diagnostic strategies.

Main Methods:

• Developed an evaluation framework using narrowband ALS (415 nm, 450 nm) and white light imaging.
• Employed skin-tone balancing, threshold-sensitivity analysis, and embedding-similarity partitioning.
• Trained models on a multi-annotator dataset with varying white-light/ALS (W/ALS) ratios.

Main Results:

• Peak accuracy was achieved with ALS-dominant training; performance declined with increased white-light images.
• Skin-tone balancing reduced failure rates for darker skin tones but increased overprediction in some subgroups.
• Image-level partitioning inflated generalization, highlighting the need for injury-level analysis.

Conclusions:

• Future research should prioritize injury-level data partitioning for more robust models.
• A weighted balance of ALS images during training is crucial for optimal performance.
• The developed framework offers improved robustness and fairness in forensic bruise detection.