Alexandros Papadopoulos, "Neural network architectures for skin cancer detection", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2025
https://doi.org/10.26233/heallink.tuc.104992
Deep learning has transformed medical imaging, particularly for automated disease diagnosis, with Convolutional Neural Networks (CNNs) traditionally dominating image analysis. The recent rise of Transformer architectures, including Vision Transformers (ViTs), offers a powerful new paradigm for image recognition. This thesis provides a systematic comparison of custom designed CNN and ViT architectures built from scratch for dermatological image classification. It explores the impact of fundamental design choices, such as network depth and regularization techniques, to maximize predictive accuracy, while considering computational constraints. These models are benchmarked on the HAM10000 dataset, a collection of over 10,000 dermatoscopic images across seven skin lesion categories, using a comprehensive suite of metrics including validation accuracy, loss, Macro F1-score, and parameter count. This study offers critical, evidence-based insights into the architectural trade-offs inherent to each approach, serving as a methodical guide for developing effective, specialized models for challenging medical imaging tasks.