Enchancing Robustness in Skin Lesion Detection: A Benchmark of 32 Models on a Novel Dataset Including Healthy Skin Images

The rising incidence of skin cancer necessitates the development of automated, reliable diagnostic tools to support clinicians. While deep learning models, particularly from the YOLO family, have shown promise, their application in real-world scenarios is limited by challenges such as class imbalance and the inability to process images of healthy skin, leading to potential false positives. This study presents a comprehensive comparative analysis of 32 object detection models, primarily from the YOLO architecture (v5–v12) and RT-DETR, to identify the most effective solution for skin lesion detection. We curated a novel, balanced dataset of 10,000 images based on the ISIC archive, comprising 10 distinct lesion classes (benign and malignant). Crucially, we introduced a dedicated ‘background’ class containing 1000 images of clear skin, a novelty designed to enhance model robustness in clinical practice. Models were systematically evaluated and filtered based on performance metrics (mAP, Recall) and complexity. Through a multi-stage evaluation, the YOLOv9c model was identified as the superior architecture, achieving a mAP@50 of 72.5% and a Recall of 71.6% across all classes. The model demonstrated strong performance, especially considering the dataset’s complexity with 10 classes and background images. Our research establishes a new benchmark for skin lesion detection. We demonstrate that including a ‘background’ class is a critical step towards creating clinically viable tools. The YOLOv9c model emerges as a powerful and efficient solution. To foster further research, our curated 10-class dataset with background images will be made publicly available.