A critical comparison between template-based and architecture-reused deep learning methods for generic 3D landmarking of anatomical structures

Abstract

Shape alterations in body organs are common pathological hallmarks of multiple disorders, making quantitative shape analysis key for obtaining diagnostic and prognostic biomarkers. In this context, Geometric Morphometrics (GM) is a powerful approach to capture subtle yet significant dysmorphologies. Since GM relies on registering landmarks on 3D anatomical structures, developing generic, automatic and accurate 3D landmarking methods is key for building high-throughput morphometric tools. This study compares state-of-the-art deep learning and template-based 3D landmarking methods using MRI datasets of faces, upper airways, and hippocampi. We evaluated these methods in terms of landmarking error and morphometric variables relative to manual annotations. Our results show that architecture-reused deep learning methods are more accurate and faster in inference than template-based techniques, particularly for anatomical structures with high shape variability, even with fewer training examples.