Main   Thesis  

Dermoscopic attribute detection

The goal of my master’s thesis project is the detection of certain structures in dermoscopic images to aid in the diagnosis of melanoma.

This is achieved through machine learning using a UNet-architecture with a compact encoder to segment the structures. To address problems with the dataset, including heavy data imbalance, two families of loss functions and an oversampling technique are evaluated. To improve interpretability, heatmaps of the model output are produced.