Wet Poster Session

In recent years, deep learning methods have received significant attention for computer-aided diagnosis in a variety of fields of medical imaging. They outperform former methods in capability and accuracy. However, deep models trained for diagnosis of specific cases currently lack the ability to say "I don't know" for ambiguous or unknown cases. Therefore, this work proposes the integration of prediction uncertainties into diagnostic classifiers to increase patient safety in deep learning. We train the ResNet-34 image classifier on a dataset of 84.484 optical coherence tomographies showing four different retinal conditions. Monte Carlo sampling is applied with dropout at test time for uncertainty estimation. Dropout with p=0.5 is added before the last fully connected layer, creating a probabilistic classifier. In Monte Carlo experiments, 100 forward passes are performed to get a posterior distribution of the class labels. The variance of the posterior is used as metric for the uncertainty. A study is performed to show if false predictions of a deep model correlate to high prediction uncertainty. Our results shown that cases in which the network predicts incorrectly correlate with a higher uncertainty. Mean uncertainty of incorrectly diagnosed cases was 8.7 times higher than mean uncertainty of correctly diagnosed cases. In addition, it was observed that a higher prevalence of a disease in the data set correlates with a lower mean uncertainty. The findings were even stronger when training the classifier with smaller data sets. Modeling of the prediction uncertainty in computer-aided diagnosis with deep learning yields more reliable results and is therefore anticipated to increase patient safety. This can help to transfer such systems into clinical routine and to increase the acceptance of physicians and patients for machine learning in diagnosis. In future work, the uncertainties can be used to further increase classification accuracy.