Segmentation of Skin Lesion Using Double U-Net Framework for Enhanced Feature Extraction

Authors Pujari Madhuri, Kunchala Supriya, Kodipyaka Sai Ganesh, B. Lakshmi Prasanna
Affiliations

Department of Computer Science and Information Technology, Institute of Aeronautical Engineering, Hyderabad, India
Е-mail madhuripujari12@gmail.com
Issue Volume 17, Year 2025, Number 5
Dates Received 18 August 2025; revised manuscript received 5 October 2025; published online 30 October 2025
Citation Pujari Madhuri, Kunchala Supriya, Kodipyaka Sai Ganesh, B. Lakshmi Prasanna, J. Nano- Electron. Phys. 17 No 5, 05034 (2025)
DOI https://doi.org/10.21272/jnep.17(5).05034
PACS Number(s) 07.05.Tp, 87.19.xj
Keywords Skin lesion segmentation, Double U-Net architecture, Semantic segmentation, Dermoscopic images.
Annotation

Skin cancer is a frequent cancer globally, and the outlook for a patient and the effectiveness of treatment depend on its prompt identification. Dermoscopic images are very essential in the accurate and automatic segmentation of skin lesions for helping clinicians diagnose skin cancer. We, in this study, propose a new semantic segmentation model based on the DoubleU-Net architecture for improving the detail and accuracy of skin lesion detection. The proposed DoubleU-Net model works by integrating two U-Net networks in sequence, where the first U-Net extracts high-level features and provides an initial segmentation map. The second U-Net refines this output by learning from the residual errors of the first network and produces a more detailed and accurate segmentation. This dual network design helps in overcoming the challenges of blurred lesion boundaries and varying lesion sizes, which are common issues in skin lesion segmentation. We evaluated the performance of our model using the publicly available ISIC(2018) dataset which contains thousands of annotated dermoscopic images. Our model evolved with the Dice coefficient and losing cross-entropy in order to deal with class imbalance, which is frequently observed in medical datasets. Experimental results show that our proposed DoubleU-Net architecture performs more effectively than baseline U-Net model when using the metrics Intersection over Union (0.81589), Dice coefficient (0.88628), and overall segmentation accuracy (0.94437).

List of References

Full text download
English version of article