AULUNet: An Adaptive Ultra-Lightweight U-Net Framework for Efficient Skin Lesion Segmentation in Resource-Constrained Environments

Md Maklachur Rahman

Abstract:

High accuracy in skin lesion segmentation is crucial for mobile and embedded medical applications, where most existing models are too computationally demanding. We
introduce AULUNet, an ultra-lightweight U-Net framework that addresses this limitation through three key innovations: Adaptive Kernel Fusion (AKF) for balanced extraction of local and global features, Residual Global Fusion (RGF) for refining bottleneck
representations with global context, and Lightweight Skip Gate (LSG) for selective skip
connection enhancement. On ISIC2017, AULUNet achieves 85.32 % mIoU and 92.08
% DSC, surpassing standard U-Net by 5.77 % and 3.47 %, respectively. On ISIC2018,
it achieves 82.44 % mIoU and 90.37 % DSC, delivering gains of 7.80 % and 4.89 %
over the baseline. Notably, our model requires only 0.029 M parameters and 0.069
GFLOPs, reducing parameter count by 99.63 % (from 7.773 M to 0.029 M) and compute by 99.50 % (from 13.758 GFLOPs to 0.069 GFLOPs), while maintaining consistent
performance across diverse lesion structures. We validate the robustness of our design
through extensive experiments and ablation studies. Our code is publicly available athttps://github.com/maklachur/AULUNet.