Use of uncertainty quantification as a surrogate for layer segmentation error in Stargardt disease retinal OCT images

Semantic segmentation methods based on deep learning techniques have transformed the analysis of many medical imaging modalities, including the extraction of retinal layers from ocular optical coherence tomography images. Despite the high accuracy of these methods, the automatic techniques are not free of labelling errors, which means that a clinician may need to engage in the time-consuming process of reviewing the outcome of the segmentation method. Given this shortcoming, having access to segmentation techniques that can provide a confidence metric associated with the output (probability class map) are desirable. In this study, the use of Monte-Carlo dropout combined with a residual U-net architecture is explored as a way to provide segmentation pixel-wise prediction maps as well as corresponding uncertainty maps. While assessing the proposed network on a dataset of subjects with a retinal pathology (Stargardt disease), the uncertainty map exhibited a high correlation with the boundary error metric. Thus, confirming the potential of the technique to extract metrics that are a surrogate of the segmentation error. While the Monte-Carlo dropout seems to have no detrimental effect on performance, the uncertainty metric derived from this technique has potential for a range of important clinical (i.e. ranking of scans to be reviewed by a human expert) and research (i.e. network fine-tuning with a focus on high uncertainty/high error regions) applications.