Model Quality in AI-based Bruise Detection: Rethinking IoU and Confidence Thresholds

Artificial Intelligence (AI)-based object detection models, like YOLO and Faster R-CNN, depend heavily on the Intersection over Union (IoU) and confidence thresholds to evaluate model performance. However, fixed thresholds can be biased and may have a disparate effect across subpopulations, even when traditional performance metrics suggest strong model performance. This paper examines how varying IoU and confidence thresholds affect standard evaluation metrics such as precision, recall, F1-score, and mean Average Precision (mAP) along with their effect on five widely used fairness metrics - Demographic Parity, Equalized Odds, Equality of Opportunity, Accuracy Equality, and Disparate Impact. This study evaluated a dataset of bruise images under natural and alternative light sources and found that fairness and performance trade-offs can be mitigated by selecting intermediate threshold values rather than fixed extremes. In addition, the results highlight the need for dynamical optimization of thresholds to achieve both, high model performance and fairness, in AI-driven bruise detection.