Uncertainty Quantification in Computer-Aided Diagnosis: Make Your Model say "I don't know" for Ambiguous Cases

verfasst von
Max-Heinrich Laves, Sontje Ihler, Tobias Ortmaier

We evaluate two different methods for the integration of predictionuncertainty into diagnostic image classifiers to increase patient safety indeep learning. In the first method, Monte Carlo sampling is applied withdropout at test time to get a posterior distribution of the class labels(Bayesian ResNet). The second method extends ResNet to a probabilistic approachby predicting the parameters of the posterior distribution and sampling thefinal result from it (Variational ResNet).The variance of the posterior is usedas metric for uncertainty.Both methods are trained on a data set of opticalcoherence tomography scans showing four different retinal conditions. Ourresults shown that cases in which the classifier predicts incorrectly correlatewith a higher uncertainty. Mean uncertainty of incorrectly diagnosed cases wasbetween 4.6 and 8.1 times higher than mean uncertainty of correctly diagnosedcases. Modeling of the prediction uncertainty in computer-aided diagnosis withdeep learning yields more reliable results and is anticipated to increasepatient safety.

Institut für Mechatronische Systeme