Integrating clinically-relevant features into skin lesion classification
To increase both the accuracy and the explainability of computer-aided diagnosis systems, a multi-task network was built to perform skin lesion classification. Clinically-relevant features were provided as secondary output during training, next to the ground truth classification.
These clinically-relevant features were either the absence/presence of certain malignancy signs within the skin lesion, or segmentation masks of them. The former did not significantly increase the performance, whereas the latter did.
After receiving a bachelor’s degree in Psychobiology, Emmeke went on to pursue a master’s degree in Artificial Intelligence, both at the University of Amsterdam. After graduating last August, she started a PhD at the Vrije Universiteit Amsterdam, on the topic of prediction of crowd dynamics.