Lightweight Skin Lesion Segmentation for Edge Deployment: A Critical Review of Architectures, Accuracy Compensation, and Clinical Translation

Abstract

[EN]Automated skin lesion segmentation via deep learning has achieved remarkable accuracy in controlled laboratory settings, yet its clinical translation to point-of-care (PoC) edge devices remains fundamentally constrained by the computational demands of state-of-the-art architectures. This review systematically bridges the gap between algorithmic innovation and hardware-aware deployment by presenting a critical analysis spanning three interconnected dimensions: 1) architectural evolution from convolutional neural networks (CNNs) through Vision Transformers to emerging linear-complexity State Space Models (e.g., Mamba), evaluated against a unified efficiency–accuracy Pareto framework rather than isolated accuracy rankings; 2) mathematical formalisation of lightweighting-induced precision degradation and its compensation via knowledge distillation, structural reparameterisation, and frequency-domain enhancement, with explicit assessment of evidence quality and reproducibility across studies; and 3) empirical deployment profiling across heterogeneous edge platforms including NVIDIA Jetson, Intel OpenVINO, and mobile processors. Through a multi-dimensional evaluation framework integrating spatial overlap (Dice), boundary fidelity (HD95), and on-device throughput (FPS), supplemented by a meta-analysis of dataset usage biases and validation protocol inconsistencies, we establish that heterogeneous multi-paradigm fusion architectures can preserve clinical-grade segmentation quality under severe resource constraints—provided that evaluation methodologies are standardised. Critically, our quantitative analysis of 36 reviewed studies reveals that approximately 65% of reported performance figures are derived from non-standardised data splits, rendering direct cross-study comparisons statistically unreliable. We further identify critical open challenges—including cross-domain generalisation, hardware–algorithm co-design, demographic fairness, and edge-native federated learning—an...