SKIN CANCER DETECTION USING DEEP LEARNING METHODS

Abstract

Skin cancer is one of the most common and dangerous types of cancer across the world, with around 1.5 million new cases being diagnosed every year. Traditional methods of diagnosis work effectively but suffer from problems related to inter-observer variability and specialist availability. This thesis presents two complementary contributions. First, a systematic review of eighteen deep learning studies published between 2020 and 2025 is conducted, covering convolutional neural networks, Vision Transformers, ensemble methods, and hybrid models on benchmark datasets including HAM10000 and the ISIC archive. Key research gaps identified include the absence of reproducible Vision Transformer benchmarks on standard dermoscopy datasets. Secondly, Low-Rank Adaptation (LoRA) is performed on ViT-B/16 for multiclass der moscopic classification on the HAM10000 dataset by fine-tuning only 1.36% of all the model’s parameters. The proposed LoRA-ViT network provides a test-set macro F1 score of 0.7222, ac curacy of 78.04%, and macro AUC-ROC of 0.9577, surpassing MobileNetV3 and approaching ResNet-50 with significantly lower computational overhead. The best choice, according to the rank ablation experiment, is r = 32. The presented experiments show that parameter-efficient fine-tuning of Vision Transformers can be considered a good method for detecting skin cancer from dermoscopic images under limited resources.