Using Out-of-Distribution Detection for Model Refinement in Cardiac Image Segmentation
Abstract
We introduce a new learning framework that builds upon the recent progress achieved by methods for quality control (QC) of image segmentation to address the poor generalisation of deep learning models in Out-of-Distribution (OoD) data. Under the assumption that the label space is consistent across data coming from different distributions, we use the information provided by a QC module as a proxy of the segmentation model's performance in unseen data. If the model's performance is poor, the QC information is used as feedback to refine the training of the segmentation model, thus adapting to the OoD data. Our method was evaluated in the context of the Multi-Disease, Multi-View & Multi-Center Right Ventricular Segmentation in Cardiac MRI Challenge reporting average Dice Score and Hausdorff distance of 0.905 and 10.472, respectively.
Origin : Files produced by the author(s)