Precise Image Segmentation using Explainable U-Net based Hybrid Approach for Improved Detection of Skin Lesion

Document Type

Article

Publication Title

International Journal of Computing and Digital Systems

Abstract

Early detection and prevention of skin cancer such as Melanoma is essential to reduce death rate across the world. This obstacle can be addressed through automated image segmentation of skin lesions from dermatoscopic or image-based data. Identification of affected area or region though Image segmentation integrated with U-Net based deep learning framework is gaining popularity in recent times. However, there is still research gaps in reliable automated image segmentation approaches for detecting skin cancer. Significant challenges are vanishing gradients problems in multi scale objects such as skin lesions, loss of important spatial information and lack of deep learning model transparency. This paper is aiming to address the above-mentioned challenge through a model hybridization approach combining U-Net based model along with ResNet and Attention mechanism for skin lesion image segmentation. To evaluate the proposed model, the publicly available PH2 dataset containing dermoscopic images of skin lesions along with their corresponding ground truth segmentation is used. Furthermore, to augment the explainability for the proposed hybridization approach, explainable Artificial Intelligence (XAI) method saliency map has been implemented. The proposed mechanism in this paper provides better comprehensibility about the active regions of the image as well as the performance with a higher Test Dice coefficient of 0.913 and least Test error of 0.199. In addition, a set of significant metrics is proposed in the paper to evaluate the performance of the proposed model and similar kinds of future models.

DOI

10.12785/ijcds/1571149929

Publication Date

1-1-2025

This document is currently not available here.

Share

COinS