Routine echocardiographic quantification can be impaired by operator variability and laborious post‑processing, and real‑time cine ultrasound requires low‑latency automated methods for clinical use. To address the lack of standardized evaluation frameworks and verified edge deployment for cardiac elastography processing, the authors developed MyoSTAT.AI, a deterministic, fully reproducible toolkit for myocardial stiffness tracking and shear‑wave velocity estimation. The system was evaluated entirely on synthetic reference corpora to provide a controlled benchmarking environment and reproducible comparisons across architectures.
MyoSTAT.AI comprises two primary branches: a segmentation pipeline and a shear‑wave elastography (SWE) velocity estimation and stabilization pipeline. The segmentation component is based on U‑Net family architectures with both single‑frame and temporal variants. The SWE branch performs propagation‑direction estimation using the Radon transform, followed by time‑domain shear‑wave speed estimation and an optional temporal stabilization module. The authors emphasize determinism and reproducibility in the pipeline design and reporting.
All experiments used synthetic data created as reference corpora. The segmentation corpus contained 1,200 synthetic frames and the SWE corpus contained 900 synthetic velocity‑field cases. Experiments and reported metrics were calculated on these synthetic evaluation sets. The authors note that the excitation source and acquisition physics were not modeled in the synthetic datasets and explicitly state that real‑data acquisition and clinical validation are required for translation.
Four U‑Net–based architectures were evaluated through systematic ablation to compare spatial and temporal modeling approaches:
The study measured segmentation accuracy using Dice and Intersection over Union (IoU) on the synthetic segmentation corpus. Temporal designs demonstrated substantially improved accuracy over the 2D baseline, with trade‑offs in latency depending on model complexity.
The SWE branch estimates propagation direction using a Radon‑transform approach and computes shear‑wave speed in the time domain from synthetic velocity fields. A configurable temporal stabilization module was implemented; the authors report results using penalized least‑squares smoothing (smoothn) with regularization parameter λ = 10. Temporal stabilization was evaluated on synthetic SWE cases, assessing effects on coefficient of variation (CoV) of shear‑wave speed and on edge preservation.
To evaluate inference feasibility for real‑time usage, the authors exported the UNet2.5D segmentation model and performed TensorRT FP16 deployment benchmarks on two hardware platforms:
Latency comparisons among segmentation architectures were reported; for example, a 2.5D model with a 3‑frame temporal window produced Dice 0.983 at a latency of 433 ms, compared with a higher‑latency model measured at 1193 ms.
On the synthetic evaluation quantities reported by the authors:
The authors explicitly state this work is a synthetic proof‑of‑concept rather than a clinically validated tool. Important limitations reported in the source include:
Next steps required before clinical deployment, as noted by the authors, are acquisition of real echocardiographic SWE data, modeling of acquisition physics, and formal clinical validation to demonstrate diagnostic robustness and operator independence.
Competing interests and ethics: the authors declared no competing interests and stated that relevant ethical guidelines were followed; any IRB or oversight details were described in the source text where applicable.