Uncertainty-Triggered Refinement for On-Device Skin Lesion Boundary Segmentation

Abstract

The global incidence of skin cancer, particularly melanoma, necessitates rapid, accurate, and accessible diagnostic tools. While deep learning models have achieved expert-level performance in automated skin lesion analysis, their deployment in clinical settings is hindered by the massive computational requirements of state-of-the-art architectures. Transitioning these models to resource-constrained edge devices, such as smartphones and portable dermoscopes, offers substantial benefits in terms of data privacy, reduced latency, and offline availability. However, lightweight models designed for on-device execution frequently struggle with precise boundary segmentation, an critical step for extracting morphological features used in dermatological diagnosis. Lesion boundaries are often obscured by artifacts such as hair, varying illumination, and low contrast with the surrounding healthy skin. To bridge the gap between efficiency and accuracy, this paper introduces an uncertainty-triggered refinement framework tailored for on-device skin lesion boundary segmentation. The proposed system utilizes a highly compressed baseline network to generate an initial coarse segmentation alongside a pixel-wise uncertainty map. Regions exhibiting high predictive entropy, typically localized along the lesion perimeter, trigger a localized, computationally inexpensive refinement module. By dynamically allocating computational resources only to challenging ambiguous regions, the framework maintains the low latency required for mobile applications while achieving segmentation accuracy comparable to heavyweight cloud-based models. Extensive evaluations on standard dermoscopic datasets demonstrate the superior trade-off between resource utilization and boundary fidelity. The results highlight the potential of uncertainty-aware dynamic computation in advancing reliable, accessible point-of-care medical image analysis