Utilizing CNNs on shear wave M-mode data to differentiate cardiac patients from healthy volunteers

Cardiac shear wave elastography (SWE) allows for the non-invasive assessment of myocardial stiffness- a key element of diastolic function - based on the speed of shear waves (SWs) that travel through the heart after valve closure. The potential of SWE has been explored in various clinical studies. However, currently, SW speed assessment mostly relies on manual or semi-automatic measurements, increasing inter- and intra-observer variability. Therefore, we investigated if an AI-based model, using a convolutional neural network (CNN), could automatically differentiate patients with diastolic dysfunction (DD) from healthy volunteers (HV) based on SW properties. 86 patients with DD and 57 health volunteers (HV) were included. SWs were visualized in M-modes of the septum, color coded for tissue acceleration. These M-modes were used as input for the CNN model as well as conventional SW speed (SWS) measurements. Where the conventional approach detecting DD based on measuring SWS resulted in an area under the curve (AUC) of 0.78, the AI-based model had an average AUC of 0.91±0.05 across all cross-validation folds. Further, there was a moderate correlation between measured SWS and the model’s predictions (r=0.4; p<0.001). A grad-CAM analysis confirmed that the model’s output was mainly determined by the shear wave propagation pattern in the M-mode. In conclusion, this study shows that an AI model can differentiate between DD patients and HV based on SW properties in a fully automated manner. Future work should focus on expanding the dataset, improving accuracy and identifying the SW properties that contribute most to accurate disease classification.