Net4DMD: A Deep Learning Framework for 4D CMR Left Ventricle Strain Analysis in Duchenne Muscular Dystrophy

Background: Cardiomyopathy is the leading cause of mortality in Duchenne Muscular Dystrophy (DMD), yet current clinical metrics often fail to detect early myocardial dysfunction. 4D cardiovascular magnetic resonance (CMR) strain analysis can capture subtle mechanical abnormalities [1][2], but its adoption is limited by the lack of accessible, standardized post-processing tools. We developed py4DCMR, a Python toolkit for left ventricle (LV) strain quantification, and Net4DMD, a collection of spatiotemporal deep learning models for automated segmentation and feature tracking in DMD CMR datasets.

Methods: py4DCMR provides an end-to-end pipeline including DICOM import, reorientation, generation of 4D endocardial and epicardial point clouds and computation of circumferential, radial and surface area strain rates. Net4DMD incorporates temporal convolution mechanisms, enabling the network to learn cardiac motion patterns specific to DMD pathology. Using Net4DMD’s predictions, py4DCMR creates a 3D mesh for each timepoint in the cardiac cycle, which are then used to generate spatiotemporal strain maps for circumferential, radial, and surface area strain rates. Model training was performed on multi-institutional CMR datasets of healthy controls (n = 10) and DMD patients (n = 81) imaged annually over 3 consecutive visits, n scans = 170. Performance was assessed using Dice scores for segmentation and tolerance-based pass rates for tracking and orientation accuracy.

Results: Each deep learning-assisted step was evaluated individually for consistency. The orientation pipeline (Figure 1.A) achieved acceptance rates of 81%, 82%, 76% and 84% for center x coordinate, center y coordinate, cropping bounds and rotation angle (threshold ± 0.050 mm for center x/y and cropping bounds, ± 15 degrees for the rotation angle) respectively. The LAX boundary tracking (Figure 1.B) pipeline achieved 72%, 68% and 22% acceptance rates for endocardial apex, epicardial apex and base displacement curves (threshold ± 0.1 mm), respectively. For the SAX segmentation step (Figure 1.C), Dice scores were 78% for the LV cavity and 31% for the LV wall.

Conclusion: Our integrated framework enables high-throughput, reproducible quantification of myocardial strain in DMD using 4D CMR. With improvements in LV base longitudinal tracking and LV wall segmentation, this approach has the potential to support longitudinal monitoring and accelerate clinical translation of advanced CMR into DMD care. Future steps include expanding the dataset with additional control and DMD images from a National Institutes of Health R01-funded project aiming to acquire 4,000 MRI scans, as well as exploring novel network architectures, particularly for LAX boundary tracking.

Figure 1: Net4DMD pipeline. Cine cardiac MRI stacks are reoriented, centered, and cropped, followed by long-axis tracking and short-axis segmentation of the left ventricle (LV). The pipeline integrates three neural network architectures: (A) a 3D U-Net encoder that regresses orientation parameters (center point coordinates, rotation angle and cropping boundaries), (B) a spatiotemporal convolutional network that models dynamic boundary (endocardial apex, epicardial apex and base) motion across the cardiac cycle, and (C) a 2D U-Net 3+ that segments the left ventricle (LV) wall in short-axis slices over time. LV masks are converted into a 48-point mesh for quantification of circumferential, radial, and surface area strain rates.