A Lightweight U-Net Model for Accurate Skin Lesion Segmentation

In this paper, a new lightweight U-Net deep learning-based neural network designed for the segmentation of skin lesions is proposed. Segmentation of skin lesions is the most critical step in computer-aided dermatology diagnosis for the early detection of melanoma and other diseases. However, we address the difficulty related to the precise definition of the lesion margins with an eye on the computation cost. We have demonstrated the state-of-the-art performance of DeepSkinSeg in most metrics on dermoscopic images using the PH2 and Human Against Machine (HAM10000) datasets. The metrics of the DeepSkinSeg model were robustness measured as the Intersection over Union (IoU) at 91.49, Dice coefficient at 95.56, precision at 97.97, sensitivity at 96.84, and accuracy at 96.71 for the PH2 dataset. Other standard generalization capabilities for the HAM10000 dataset could be an IoU of 92.97, a Dice coefficient of 96.36, precision at 97.64, sensitivity at 95.10, and an accuracy of 94.59. DeepSkinSeg has a very efficient inference because the model itself is lightweight, proving to be very helpful for real-time dermatological analysis. This work further advanced the computer-aided diagnosis in the task of skin lesion classification, guaranteeing even more promising clinical applications.