Deformable Image Registration for Self-supervised Cardiac Phase Detection in Multi-View Multi-Disease Cardiac Magnetic Resonance Images

TL;DR

The paper addresses temporal misalignment in cine CMR by proposing a fully self-supervised deformable registration framework that yields a dense vector field $\phi_t$ and a one-dimensional motion descriptor $\alpha_t$, enabling detection of five keyframes: $ed$, $es$, $ms$, $pf$, and $md$. This approach operates in SAX and 4CH views and relies on a focus-point–based masking strategy, avoiding ECG labels. Evaluated on public multi-center datasets (MM2, MNMS, ACDC, GCN), it achieves substantial improvements in cyclic frame difference for $ed$ and $es$ compared with a volume-based baseline, with mean $cFD$ under 1.31 frames in SAX and 1.73 frames in LAX, and demonstrates generalization to rare congenital conditions. The work facilitates temporally aligned inter- and intra-patient cardiac dynamics analyses and supports extended phenotyping, with code released for reproducibility and further research.

Abstract

Cardiovascular magnetic resonance (CMR) is the gold standard for assessing cardiac function, but individual cardiac cycles complicate automatic temporal comparison or 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 long-axis (4CH) cine CMR. Initially, dense deformable registration fields are derived from the images and used to compute a 1D motion descriptor, which provides valuable insights into global cardiac contraction and relaxation patterns. From these characteristic curves, keyframes are determined using a simple set of rules. The method was independently evaluated for both views using three public, multicentre, multidisease datasets. M&Ms-2 (n=360) dataset was used for training and evaluation, and M&Ms (n=345) and ACDC (n=100) datasets for repeatability control. Furthermore, generalisability to patients with rare congenital heart defects was tested using the German Competence Network (GCN) dataset. Our self-supervised approach achieved improved detection accuracy by 30% - 51% for SAX and 11% - 47% for 4CH in ED and ES, as measured by cyclic frame difference (cFD), compared with the volume-based approach. We can detect ED and ES, as well as three additional keyframes throughout the cardiac cycle with a mean cFD below 1.31 frames for SAX and 1.73 for LAX. Our approach enables temporally aligned inter- and intra-patient analysis of cardiac dynamics, irrespective of cycle or phase lengths. GitHub repository: https://github.com/Cardio-AI/cmr-multi-view-phase-detection.git

Figures (15)

• Figure 1: Graphical abstract. Overview of the proposed pipeline. The top row illustrates the input data and the self-supervised deformable image registration model. The bottom row shows the interpretation of the resulting dense deformable vector field as motion direction, enabling derivation of a one-dimensional motion descriptor for cardiac key-frame detection.
• Figure 2: LV volume curve, ECG signal and our proposed motion descriptor $\alpha$ over the cardiac cycle. The figure shows the temporal relation between cardiac phases and keyframes, left ventricular volume (top blue curve), ecg (middle red curve), and the motion descriptor $\alpha$ (bottom black curve) derived from cmr data. The cycle is divided into systole (blue) and diastole (red), with iso-volumetric contraction time (ICT) and iso-volumetric relaxation time (IRT) in respectively lighter shades. The bottom curve depicts $\alpha$, where a negative value for $\alpha$ indicates contractile motion, and positive values refer to relaxing cardiac motion. Characteristic points in $\alpha$ align with physiological events (Section \ref{['sec2_3:cardiac_key_frames']}).
• Figure 3: Self-supervised rule-based masking of cmra) Schematic illustration of computation of the direction of motion $\alpha$. The motion vector $\vec{v}$ from each voxel $\mathbf{x}_i$ is compared to a reference position vector $\vec{w}$, which points from the corresponding voxel to a fixed anatomical focus point $C_n$. The angle between $\vec{v}$ and $\vec{w}$ is quantified by their cosine similarity $\alpha = \cos(\vec{v}, \vec{w}) \in [-1,1]$. This scalar $\alpha$ represents the directional relationship: negative values below $0$ indicate contraction (motion toward $C$), while positive values indicate relaxation (motion away from $C_n$). b & c) The first row shows the original cmr slice at a single time point from either the 4ch (b) or sax (c) view. The grid below presents filtered directional motion fields $\alpha$ for the same frame, visualized at varying thresholds. Columns correspond to increasing directional change thresholds $T_{\Delta\alpha}$, and rows to increasing motion magnitude percentiles $T_{\text{norm}}$. Blue indicates contractile motion ($-1 \leq \alpha < 0$) directed toward the focus point $C$, and red indicates relaxing motion ($0 < \alpha \leq 1$) away from it. The top-left cell ($T_{\Delta\alpha} = 0.0$, $T_{\text{norm}} = 0$) shows the raw, unfiltered deformation field.
• Figure 4: Summary of datasets used, including patient groups (see Table \ref{['tab:pathology']}), segmentation annotations, usage, imaging views (4ch: four-chamber, sax: short-axis), number of cases ("Num. cases") and keyframe annotations ("Keyframes") per view. Segmentation annotations are published bi-ventricular segmentation at ed and es. "Keyframes" include either the original dataset annotations or additional physician annotations (asterisk*; Section \ref{['sec2_3:cardiac_key_frames']}). "All" denotes all five keyframes (ed, ms, es, pf, md). Values in "Num. cases" and "Keyframes" match the order of listed views.
• Figure 5: Cyclic frame difference (mean ± SD) for the sax view for five datasets (mnms, mm2 test and training, acdc, gcn) with respect to different focus points $C_n$. $base$ - supervised volume-based approach, $C_{mse}$ and $C_{vol}$ are computed fully self-supervised without prior anatomical knowledge, while $C_{sept}$ and $C_{lv}$ uses the segmentation model to compute anatomical focus point. Best results are marked in bold. The iov is calculated between public annotations and our annotations. The first row of the mnms dataset include the results reported by garcia2023cardiac. $\ast: p < 0.05$, $\ast\ast: p < 0.01$ (vs. base, paired Wilcoxon test). NR - not reported.