Interpretability of Uncertainty: Exploring Cortical Lesion Segmentation in A number of Sclerosis
Authors: Nataliia Molchanova, Alessandro Cagol, Pedro M. Gordaliza, Mario Ocampo-Pineda, Po-Jui Lu, Matthias Weigel, Xinjie Chen, Adrien Depeursinge, Cristina Granziera, Henning Müller, Meritxell Bach Cuadra
Summary: Uncertainty quantification (UQ) has grow to be vital for evaluating the reliability of synthetic intelligence programs, particularly in medical picture segmentation. This research addresses the interpretability of instance-wise uncertainty values in deep studying fashions for focal lesion segmentation in magnetic resonance imaging, particularly cortical lesion (CL) segmentation in a number of sclerosis. CL segmentation presents a number of challenges, together with the complexity of guide segmentation, excessive variability in annotation, knowledge shortage, and sophistication imbalance, all of which contribute to aleatoric and epistemic uncertainty. We discover how UQ can be utilized not solely to evaluate prediction reliability but in addition to offer insights into mannequin habits, detect biases, and confirm the accuracy of UQ strategies. Our analysis demonstrates the potential of instance-wise uncertainty values to supply publish hoc world mannequin explanations, serving as a sanity examine for the mannequin. The implementation is offered at https://github.com/NataliiaMolch/interpret-lesion-unc. △
