OPTIMIZED LIGHTWEIGHT DUALENCODER FRAMEWORK WITH KNOWLEDGE DISTILLATION FOR REALTIME SKIN LESION SEGMENTATION

Project Code :TCMAPY2321

Objective

The objective of this project is to develop an optimized lightweight Dual Encoder framework for real-time skin lesion segmentation and multi-class skin disease classification. It integrates U-Net+ and YOLOv8 for segmentation and detection, while a dual encoder combining ResNet50 and EfficientNetB0, along with MobileNet and CNN models, improves classification accuracy. Knowledge distillation ensures efficient, interpretable, and scalable deployment with Grad-CAM visualization.

Abstract

Skin diseases have become a significant health concern globally, affecting millions of individuals across different age groups. Early detection and accurate diagnosis are critical for effective treatment and management of skin disorders. This research focuses on the development of an optimized lightweight DualEncoder framework for skin lesion segmentation and multi-class skin disease classification. The framework leverages advanced machine learning techniques, including U-Net+ for lesion segmentation and YOLOv8 for real-time detection, alongside CNN, MobileNet, and ResNet for the classification of various skin conditions. A dual encoder architecture is employed, combining the strengths of two pre-trained models, ResNet50 and EfficientNetB0, to enhance the accuracy and robustness of the system. The primary goal is to deliver a system that efficiently segments skin lesions and classifies them into categories such as Melanoma, Eczema, BCC (Basal Cell Carcinoma), and others. The system is designed with a user-friendly interface, powered by Flask in the backend and HTML, CSS, and JS for the frontend, ensuring accessibility across platforms. By applying knowledge distillation techniques, the system can perform segmentation and classification tasks with reduced computational cost, making it suitable for deployment in environments with limited resources. To enhance model transparency, Grad-CAM is used to visualize the regions of the skin images that contribute most to the model’s predictions, providing interpretable insights into the decision-making process. The results indicate promising performance in both lesion segmentation and disease classification, achieving high accuracy rates and demonstrating the potential of deep learning-based models in the field of dermatology.

Keywords: Skin disease detection, skin lesion segmentation, dual encoder, U-Net+, YOLOv8, ResNet50, EfficientNetB0, deep learning, knowledge distillation, Grad-CAM

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