Hybrid Wavelet-Swin Transformer and Dual-Path Attention Networks for Robust Skin Lesion Segmentation

Project Code :TCMAPY2480

Objective

Hybrid Wavelet-Swin Transformer and Dual-Path Attention Networks for Robust Skin Lesion Segmentation project presents an advanced deep learning framework for accurate skin lesion segmentation using dermoscopic images from the ISIC 2017 dataset. The proposed system aims to improve lesion boundary detection and feature extraction by combining convolutional and transformer-based architectures. Two segmentation models are developed and evaluated: WSB-Net, which integrates Wavelet Transform, Swin Transformer, and Boundary Refinement modules to capture fine-grained texture details and precise lesion boundaries, and DCT-Net, which combines a Dual-Path CNN, Transformer encoder, and Cross-Attention mechanism to learn both local and global image representations. A Flask-based web application is implemented to provide an interactive platform where users can register, log in, upload skin lesion images, and obtain segmented outputs. By leveraging complementary feature learning strategies, the framework enhances segmentation accuracy, boundary precision, and overall model robustness. The project demonstrates the effectiveness of hybrid deep learning techniques for skin lesion analysis and provides a practical solution for automated lesion segmentation.

Abstract

Skin lesion segmentation plays an important role in the analysis of dermoscopic images and supports the identification of different skin abnormalities. Accurate segmentation helps in isolating lesion regions from surrounding healthy skin, which improves the quality of further analysis. This project presents a hybrid deep learning framework titled "Hybrid Wavelet-Swin Transformer and Dual-Path Attention Networks for Robust Skin Lesion Segmentation." The proposed system utilizes the ISIC 2017 skin lesion dataset for training and evaluation. Two advanced segmentation models are implemented and compared. The first model, WSB-Net, combines Wavelet Transform, Swin Transformer, and Boundary Refinement techniques to capture detailed texture information and produce accurate lesion boundaries. The second model, DCT-Net, integrates a Dual-Path CNN, Transformer architecture, and Cross-Attention mechanism to learn both local and global image features effectively. The system is developed using Flask with an interactive web interface that includes modules such as Home, Register, Login, Segmentation, and Logout. Users can upload dermoscopic images and obtain segmented lesion outputs through the application. The framework focuses on improving segmentation quality, boundary precision, and feature representation while maintaining an efficient workflow. Experimental evaluation demonstrates the effectiveness of combining convolutional and transformer-based architectures for skin lesion segmentation tasks. The developed system provides a reliable platform for analyzing skin lesion images using modern deep learning approaches.

Keywords: Skin Lesion Segmentation, Deep Learning, Swin Transformer, Wavelet Transform, Cross-Attention, Convolutional Neural Network, Image Segmentation, Dermoscopic Images, Boundary Refinement, Flask Framework

NOTE: Without the concern of our team, please don't submit to the college. This Abstract varies based on student requirements.