Detecting When Pre-trained nnU-Net Models Fail Silently for Covid-19 Lung Lesion Segmentation

Task-agnostic Continual Hippocampus Segmentation for Smooth Population Shifts

“…Inspired by previous research on OOD detection for semantic segmentation [9], we detect data shifts by calculating the Mahalanobis distance D M (z; µ, Σ) to the training distribution. In contrast to other methods for assessing similarity, such as the Gram distance popular in rehearsal-based continual learning [21,22], the Mahalanobis distance requires storing only µ and Σ.…”

“…The history of the new model B i is initialized with B, so the history of each model contains information pertaining to all data distributions used to train it. Following previous research [9] we normalize the distances between the minimum and doubled maximum in-distribution values, and set ξ = 2µ.…”

A Dempster-Shafer approach to trustworthy AI with application to fetal brain MRI segmentation

“…Artificial Intelligence (AI) algorithms for medical image segmentation can reach super-human accuracy on average [28] and yet most radiologists do not trust them [3,5]. This is partly because, for some cases, AI algorithms fail spectacularly with errors that violate expert knowledge about the segmentation task when the AI was applied across imaging protocol and anatomical pathologies [3,17,22] (Fig. 1b).…”

Distance-based detection of out-of-distribution silent failures for Covid-19 lung lesion segmentation

“…A large distance signals that the model has not seen specific activation patterns in the past, and therefore outputs produced from such novel features cannot be trusted . Our method ( Gonzalez et al, 2021 ), initially presented at MICCAI 2021, is lightweight and requires no changes to the network architecture of the training procedure, allowing it to integrate into complex segmentation pipelines seamlessly. Further, as the distance estimation process follows after training, it can provide clinically-relevant uncertainty scores for pre-trained models.…”