Myocardial Stiffness Tracking & Assessment Toolkit Driven by Artificial Intelligence For Reproducible Benchmarking of Temporal Segmentation and Shear-Wave Velocity Stabilization on Synthetic Data

Propose: Routine echocardiographic assessment can be affected by operator variability and time-consuming post-processing, which may limit rapid quantitative analysis. Real-time cine ultrasound demands low-latency automated processing for clinical usage. Existing approaches lack standardized evaluation frameworks, verified edge deployment, and structured reproducibility guarantees. To address these limitations, we built Myocardial Stiffness Tracking & Assessment Toolkit driven by Artificial Intelligence (MyoSTAT.AI). We report a deterministic, fully reproducible pipeline that performs real-time segmentation and SWE velocity estimation, evaluated entirely on synthetic data. Approach: Four U-Net-based architectures spanning single-frame and temporal designs were evaluated via systematic ablation: a 2D baseline, a 2.5D stacked-frame model, a 3D volumetric model, and a ConvLSTM variant. Experiments used synthetic reference corpora with split accounting the segmentation corpus contained 1,200 synthetic frames, and the SWE corpus contained 900 synthetic velocity-field cases. The SWE branch used Radon-transform-based propagation-direction estimation, time-domain shear-wave speed estimation, and configurable temporal stabilization on synthetic velocity fields. TensorRT FP16 deployment benchmarks were performed separately for the exported UNet2.5D segmentation model on RTX 3060 and Jetson Orin Nano hardware. Results: On the synthetic evaluation quantities, the ConvLSTM variant achieved the highest segmentation accuracy (Dice 0.994, IoU 0.987), an 18.6% improvement over the 2D baseline (Dice 0.808). A 2.5D model with a 3-frame temporal window achieved Dice 0.983 at substantially lower latency (433 ms vs. 1193 ms). TensorRT FP16 deployment yielded 341 FPS on the RTX 3060 and 90.4 FPS on the Jetson Orin Nano. Penalized least-squares temporal stabilization (smoothn, {lambda}=10) reduced shear-wave speed CoV by 63.8% at 2.7 ms latency overhead per frame, with no loss of edge preservation. Conclusions: We demonstrate a deterministic, reproducible computational benchmarking framework for cardiac segmentation and SWE velocity-field stabilization, evaluated entirely on synthetic data, together with preliminary inference feasibility on workstation and embedded hardware. These results represent a synthetic proof-of-concept rather than a validated clinical or operator-independent tool; acquisition of real echocardiographic data and clinical validation are required before any claim of diagnostic or deployment readiness can be made.