Deep Coordinate Regression for Weakly Supervised Segmentation of the Locus Coeruleus in MRI

We propose a coordinate regression driven weakly supervised segmentation approach comprising a novel Mahalanobis distance based loss function and a unimodality facilitating regularization term. We furthermore explore its use for the segmentation of difficult structures such as the locus coeruleus (LC) that are characterized by low inter-rater agreements of delineations due to substantial rater bias. The LC has been attracting increasing interest due to its role in the pathogenesis of neurodegenerative diseases such as Parkinson's Disease (PD) and Alzheimer's Disease (AD), but its in vivo analysis using Magnetic Resonance Imaging (MRI) is challenging. Thorough evaluation yielded promising results. Although the mask similarity is lower compared to a fully supervised approach, the intra-class correlation (ICC) of the extracted features suggests a good agreement.