Fig. 2.

Demonstration of metric behavior for a trained segmentation model in the context of different medical imaging modalities. The figure is showing the differences between metrics based on distance like AHD, with true negatives like Accuracy, and without true negatives like DSC. Each subplot illustrates a violin plot which visualizes the resulting scoring distribution of all testing samples for the corresponding metric and modality. For visualization purposes, AHD was clipped to a maximum of 250 (affected number of samples per dataset: dermoscopy 2.0%, endoscopy 0.3%, fundus 0.0%, microscopy 0.0%, radiology 0.5%, and ultrasound 2.5%)