Automated error localisation and correction techniques for deep-learning-based segmentation of 3D MRI sequences based on feature-derived-region aggregation

Automatic segmentation using convolutional neural networks (CNNs) has become a key tool in musculoskeletal imaging, offering substantial reductions in processing time. However, concerns about reliability often necessitate manual inspection and correction. We present a method that leverages network-derived uncertainty to automatically identify and localise segmentation errors, reducing the need for exhaustive manual review. A 3D nnU-Net was trained on delayed gadolinium-enhanced MRI of hip cartilage. Voxel-wise uncertainty scores, computed from the SoftMax outputs of ensembled sub-networks, were aggregated over feature-based supervoxels. Each region was then evaluated for its potential impact on clinically relevant metrics, generating sensitivity scores. A logistic model combined these with uncertainty data to assign risk scores, guiding attention to areas most likely to affect clinical metrics during the initial correction steps. Using these risk scores, guided supervoxel correction of just 50 supervoxels reduced the mean absolute relative error by 2.1-fold. Guided manual correction within these regions achieved a 3.5-fold reduction, an approximate 62% supervoxel correction efficiency. Correcting the top 10 regions yielded up to 88% efficiency. This approach serves as a proof-of-concept for targeted correction in hip MRI, enhancing the clinical utility of CNN-based segmentation by demonstrating that 3D feature-derived uncertainty aggregation has the potential to reduce correction burden compared to traditional 2D methods.