Thesis on CNNs for Melanoma Classification by Miguel Marinho

Summary

Miguel built the MAR-MELA-CNN Fusion Ensemble across 8,000+ skin-lesion images, optimizing for F-beta to reduce false negatives in melanoma classification.

Problem

Early melanoma detection is difficult because harmful lesions can be visually similar to benign cases, making false negatives especially costly.

My role

Miguel designed the dataset strategy, CNN fusion ensemble, model evaluation, and real-time SMS diagnosis workflow for his MSc dissertation.

Architecture

• Eleven public skin-lesion datasets were merged into a broader melanoma classification corpus.
• Six ImageNet-pretrained CNN backbones were combined into the MAR-MELA-CNN Fusion Ensemble.
• Training prioritized F-beta to reduce false negatives rather than optimizing only for raw accuracy.
• A Twilio workflow demonstrated real-time SMS diagnosis from mobile-submitted lesion images.

Outcomes

• The thesis was graded 19/20.
• The model reached an 85% F2 score and 93% AUC in the reported evaluation.
• The project connected applied deep learning research with an accessible diagnosis prototype.