Designing a Classification Algorithm for Skin Lesions

To diagnose dangerous skin lesions early, this project's research objective was to develop an algorithm to diagnose lesions into six non-overlapping classes, i.e., healthy, psoriasis, acne, eczema, benign melanoma, and malignant melanoma, and maximize the diagnosis accuracy. The algorithm developed includes techniques such as ensemble learning, feature extraction transfer learning, fine-tuning transfer learning, data augmentation and a multi-level architecture. The algorithm proposed produces 96.67% accuracy, showing the techniques were beneficial.