Deformable image registration for self-supervised cardiac phase detection in cardiac magnetic resonance images of patients with various diseases.

Abstract

Cardiovascular magnetic resonance (CMR) is widely used to assess cardiac function, but individual cardiac cycles complicate automatic temporal comparison and sub-phase analysis. Accurate cardiac keyframe detection can eliminate this problem. However, automatic methods solely derive end-systole (ES) and end-diastole (ED) frames from left ventricular volume curves, which do not provide a deeper insight into myocardial motion. We propose a self-supervised deep learning method detecting five keyframes in short-axis (SAX) and four-chamber (4CH) cine CMR. Initially, dense deformable registration fields are derived from CMR to compute a 1D motion descriptor encoding global cardiac contraction and relaxation patterns. Keyframes are derived from these characteristic curves with a set of rules. The method was independently evaluated for both views using four databases encompassing multiple centre, vendor and disease. M&Ms-2 (n=360) was used for training and evaluation; M&Ms (n=345) and ACDC (n=100) for repeatability control. Generalisability to patients with rare congenital heart defects was tested using the German Competence Network (GCN) database. A disease-stratified analysis confirmed stable performance across cardiomyopathies and congenital abnormalities. Our method improved detection accuracy by 49%/59% for SAX and 31%/39% for 4CH in ED/ES over the volume-based approach, with mean cyclic frame difference (cFD) below 1.3 and 1.2 frames for SAX and 4CH respectively. Our framework enables temporally aligned inter- and intra-patient analysis of cardiac dynamics, irrespective of cycle or phase lengths for aligned strain analysis or temporal normalisation. Code and annotations are available at: https://github.com/Cardio-AI/cmr-multi-view-phase-detection.git.