Volumetric Body Composition Through Cross-Domain Consistency Training for Unsupervised Domain Adaptation

Abstract

Computed tomography (CT) scans of the abdomen have emerged as a robust, precise, and dependable means of determining body composition. The accurate prediction of skeletal muscle volume (SMV) using slices of CT scans holds critical importance in facilitating subsequent diagnosis and prognosis. A significant proportion of research in the field of abdominal image analysis is primarily focused on the third lumbar spine vertebra (L3), owing to two prominent factors. Firstly, L3 is a large vertebra situated in the middle of the lumbar spine, rendering it less susceptible to degenerative changes in comparison to other lumbar vertebrae, making it a stable landmark. Secondly, the slice labeling in a CT volume is an intricate and time-consuming process, demanding significant human efforts, whereas labeling a single slice from a specific vertebral level is comparatively simpler. This study leverages labeled L3 slices i.e., source domain to reliably predict unlabeled lumbar region slices other than L3 i.e., target domain. We use Cross-Domain Consistency Training (CDCT) to extend network’s current knowledge, acquired through segmenting a source domain, by learning to label a target domain. A consistency is enforced between the predictions from two segmentation networks with identical lightweight architecture but have different weight initialization points. The training objective consists of supervised loss terms for the source domain data and unsupervised loss terms for the target domain data. Remarkably, our trained network exhibits a marked enhancement in performance when applied to the target domain, indicating domain invariant feature learning through cross-domain consistency training could significantly enhance a network’s generalization capability.