Scar-Aware Late Mechanical Activation Detection Network For Cardiac Resynchronization Therapy (RF_FR_336)

Background: Accurate identification of late mechanical activation (LMA) regions is crucial for effective cardiac resynchronization therapy (CRT) lead implantation planning [1,2 ,3]. Despite the impact of myocardial scar on mechanical activation in heart failure patients, current LMA detection methods from cardiac magnetic resonance (CMR) images do not account for scar tissue [4,5,6,7]. As a result, LMA is often mistakenly identified in scarred regions, which also stretch early in the cardiac cycle due to increased left ventricular pressure from healthy tissue contraction. This limitation leads to reduced LMA estimation accuracy, potentially affecting clinical decision-making in CRT planning.

Methods: We propose a scar-aware LMA detection network that simultaneously detects myocardial scar and prevents LMA localization in scarred regions. Specifically, our model integrates a myocardial scar segmentation network, using late gadolinium enhancement (LGE) CMRs [8], with an LMA detection network, using displacement encoding with stimulated echoes (DENSE) CMRs [7]. We first train the scar segmentation network and convert the predicted scar maps into vectors representing scarring ratios (e.g., the ratio of scarred pixels to total myocardium pixels) within each evenly divided myocardial sector. These vectors serve as weighting parameters in the LMA detection network, encouraging zero predictions for Time to the Onset of circumferential Shortening (TOS) in scarred regions. This helps the model focus more accurately on LMA regions, minimizing the influence of scar-induced motion abnormalities. Note that if LGE CMR images are unavailable or unnecessary, the model defaults to standard LMA detection [8]. We pre-train the scar segmentation network on a dataset of LGE-CMR images with manually annotated myocardial scars, with a 70%:15%:15% data split for training, validation, and testing. We then train the LMA detection network on 342 short-axis DENSE CMR slices, using predicted scar maps from the pre-trained segmentation model. Among these, 33 slices with myocardial scarring (with over 50% transmurality) have corresponding LGE images, 13 of which are used for testing. We compare our approach with an existing LMA detection method that does not incorporate scar information, and evaluate its performance using the mean absolute error between predicted and ground truth TOS curves labeled on DENSE strain.

Results: Fig. 2 demonstrates that our model outperforms the baseline in predicting TOS in both scarred (2.36 (0.98, 3.18)ms vs. 4.60 (3.49, 8.07)ms) and overall cases (2.87 (1.72, 4.08)ms vs. 3.46 (2.27, 4.57)ms). Fig. 3 visualizes 3D LMA maps using predicted TOS from our model and the baseline, compared to manually labeled TOS. The results show that our model successfully identifies LMA (mid-anterior) and distinguishes it from scarring (basal inferolateral), where the baseline fails.

Conclusion: This is the first deep learning network that integrates myocardial scar information directly into LMA detection. Experimental results demonstrate that our approach substantially improves the performance of LMA detection, particularly in challenging cases with myocardial scarring. This new technique has the potential to benefit CRT response rates and patient outcomes. Moving forward, we will explore the impact of scars with varying degrees of transmurality. To maximize the impact of our model, we are committed to making our algorithm and code publicly available for research and clinical practice.