Instance Segmentation in Medical Imaging: A Comparative Study of CNN and Transformer-Based Models in a Teledermatology Study-Case
Rodrigo P. S. Ribeiro, Aldo von Wangenheim · 2025
The rapid evolution of instance segmentation models necessitates empirical comparisons to guide their adoption in critical domains like medical imaging. This study evaluates four state-of-the-art architectures—Mask R-CNN, Mask2Former, YOLOv11, and YOLOv12 on a teledermatological dataset annotated for compliance-driven segmentation of rulers and patient information tags. Results demonstrate that transformer-based and hybrid models (Mask2Former, YOLOv11) significantly outperform traditional CNNs in precision-driven metrics (AP75), highlighting their suitability for medical applications. This work provides actionable insights for model selection in healthcare, emphasizing the balance between accuracy and computational efficiency.