Undergraduate Thesis

Dual Network Attention-Based Skin Cancer Classification: Leveraging Supervised and Unsupervised Segmentation with Clinical Data Fusion

• Proposed the Deep-UNet model, integrating depthwise separable convolutions, skip attention, and Atrous Spatial Pyramid Pooling (ASPP) for improved segmentation performance.

• Developed a dual-network model integrating segmented images and clinical metadata using a multi-head cross-attention mechanism to enhance classification accuracy.

• Implemented unsupervised segmentation techniques (K-Means, GrabCut) when segmentation masks were unavailable, ensuring robust feature extraction from raw images.

Supervised by: Dr. Shaikh Anowarul Fattah, Professor, Department of EEE, BUET 

Skin cancer is one of the most prevalent and rapidly increasing types of cancer, and its early detection is crucial for improving survival rates. This thesis presents a novel approach to skin cancer classification that integrates both supervised and unsupervised seg mentation techniques with a deep learning framework to enhance lesion detection accuracy. Supervised models such as U-Net, U-Net++, and DeepLabV3+ are used for precise lesion localization when annotated data is available. In the absence of such data, unsupervised methods, including K-Means clustering and GrabCut, are employed to segment lesions effectively without requiring manual annotations, making the system adaptable to various datasets. 

For the classification task, a dual-network deep learning model is proposed, which processes both original and segmented images using DenseNet201 encoders. The architecture incorporates a multi-head cross-attention mechanism to merge global and lesion-specific features, improving the model’s focus on relevant areas of the image. Additionally, clinical data such as patient demographics and lesion localization are fused with the image features to enhance the robustness and decision-making process of the model. This hybrid approach enables more accurate and context-aware skin cancer classification. 

The proposed model significantly outperforms traditional methods in terms of classification accuracy and lesion detection performance. Furthermore, the Deep-UNet architecture developed in this research offers high memory efficiency, reducing computational costs compared to other models like U-Net++ and DeepLabV3+. This makes it a promising solution for real-time deployment in clinical settings with limited computational resources. Overall, this work contributes a comprehensive, scalable framework for automated skin cancer detection that combines state-of-the-art segmentation methods and clinical data in tegration, offering a valuable tool to assist dermatologists in early diagnosis and improving patient outcomes.