LUNG LOBE SEGMENTATION WITH AUTOMATED QUALITY ASSURANCE USING DEEP CONVOLUTIONAL NEURAL NETWORKS

Despite good performance for medical image segmentation, deep convolutional neural networks (CNNs) have not been widely accepted in clinical practice as they are complex and tend to fail silently. Additionally, uncertainty in their predictions are not well understood, making them obscure and challenging to interpret. Automatically detecting possible failures in network predictions is important, as we can refer such cases for manual inspection or correction by human observers. In this paper, we analyse the uncertainty for deep CNN-based lung lobe segmentation in computed tomography (CT) scans by proposing a test-time augmentation-based aleatoric uncertainty measure. Through this analysis, we produce spatial uncertainty maps, from which a clinician can observe where and why a system thinks it is failing, and quantify the image-level prediction of failure. Our results show that such an uncertainty measure is highly correlated to segmentation accuracy and therefore presents an inherent measure of segmentation quality.