MPTN: A video-based multi-point tracking network for atrioventricular junction detection and tracking in cardiovascular magnetic resonance imaging

Background and Objective

To develop an end-to-end artificial intelligence solution—video-based Multi-Point Tracking Network (MPTN), for detecting and tracking atrioventricular junction (AVJ) points from cardiovascular magnetic resonance and deriving AVJ motion parameters.

Methods

The MPTN model consists of two modules: AVJ point detection and AVJ motion tracking. The detection module utilizes convolutional-based feature extraction and elastic regression to detect all candidate AVJ points. The tracking module employs the optimized DeepSORT algorithm to dynamically capture spatio-temporal continuity between cardiac frames. The model was trained and evaluated on datasets from 286 subjects, including normal controls and patients with heart failure, acute myocardial infarction, pulmonary arterial hypertension, and repaired tetralogy of Fallot. AVJ motion parameters, including systolic velocity S’, early diastolic velocity E’, late diastolic velocity A’, and displacements, were derived from tracked trajectories.

Results

Our MPTN model demonstrated promising performance compared to ground truth, with correlations of 0.92 for S’, 0.93 for E’, 0.89 for A’ in mitral annular motion velocities, and 0.93 for mitral annular plane systolic excursion. For tricuspid annular motion, the correlations were 0.91 for S’, 0.90 for E’, 0.87 for A’, and 0.86 for tricuspid annular plane systolic excursion. The MPTN-derived AVJ motion parameters exhibited strong diagnostic capabilities in detecting echocardiography-derived ventricular systolic and diastolic dysfunction, with an area under the curve ranging from 0.83 to 0.88 and accuracies ranging from 78 % to 85 %.

Conclusions

Our work provides an initial framework for cardiac motion tracking and function evaluation, which may support future advances in diagnosis of heart diseases.