Skin Lesion Classification
Medical Imaging & Computer Vision
This project focuses on automated medical diagnostics through the binary classification of skin lesions, specifically distinguishing between benign and malignant cases.
Deep Learning Pipeline
-
01
Segmentation:
Leveraged the SAM (Segment Anything Model) for high-precision region-of-interest extraction, ensuring the classifier focuses solely on relevant dermoscopic features.
-
02
Architecture Exploration:
Conducted comparative experiments using state-of-the-art CNNs and Vision Transformers (ViTs) to optimize diagnostic accuracy.
-
03
Generative Augmentation:
Addressed dataset imbalances by generating high-fidelity synthetic training data using StyleGAN.
By combining foundational medical imaging techniques with generative models like StyleGAN, the project achieved robust performance even with limited authentic malignant samples.