Skin Lesion Segmentation with Improved U-Net

Overview

Applied an improved U-Net architecture to the ISIC (International Skin Imaging Collaboration) dataset for automated melanoma detection, achieving a minimum Dice score of 0.8 for lesion segmentation.

Approach

• Implemented improved U-Net with encoder-decoder architecture and skip connections for precise lesion boundary delineation
• Trained on the ISIC skin lesion dataset containing annotated skin patch images with binary segmentation masks
• Optimized with Dice loss function to handle class imbalance between lesion and non-lesion pixels

Key Learnings

• Pattern recognition via deep convolutional neural networks
• Practical experience with both TensorFlow and PyTorch frameworks
• Medical image preprocessing: normalization, augmentation, and handling of varied image dimensions
• Algorithmic design with software engineering principles for reproducible research