Self-supervised Motion Descriptor for Cardiac Phase Detection in 4D CMR Based on Discrete Vector Field Estimations

Cardiac magnetic resonance (CMR) sequences visualise the cardiac function voxel-wise over time. Simultaneously, deep learning-based deformable image registration is able to estimate discrete vector fields which warp one time step of a CMR sequence following in self-supervised manner. However, despite rich source information included these 3D+t fields, standardised interpretation challenging and clinical applications remain limited so far. In this work, we show how efficiently use field describe underlying dynamic process cycle form derived 1D motion descriptor. Additionally, based on expected cardiovascular physiological properties contracting or relaxing ventricle, define set rules that enables identification five phases including end-systole (ES) end-diastole (ED) without usage labels. We evaluate plausibility descriptor two multi-disease, -center, -scanner short-axis datasets. First, by reporting quantitative measures such as periodic frame difference for extracted phases. Second, comparing qualitatively general pattern when temporally resample align descriptors all instances across both The average ED, ES key our approach $$0.80\pm {0.85}$$ , $$0.69\pm {0.79}$$ slightly better than inter-observer variability ( $$1.07\pm {0.86}$$ $$0.91\pm {1.6}$$ ) supervised baseline method $$1.18\pm {1.91}$$ $$1.21\pm {1.78}$$ ). Code labels are available GitHub repository. https://github.com/Cardio-AI/cmr-phase-detection .