Characterization of fibrotic liver tissue microstructure for predicting shear wave speed variability: a machine-learning-based computational study

<i>Objective.</i>This study aimed to establish a link between the microstructure of simulated fibrotic liver tissues and the measured shear wave speed (SWS) variability using a machine-learning (ML)-based approach.<i>Approach</i>. Fibrotic liver tissues were simulated using biphasic random fields. The underlying microstructure of the simulated fibrotic liver pathology (sFLP) was characterized using spatial pattern distribution analysis. A ML technique was implemented to identify top-rated spatial characteristic (SC) features and provide context for SWS variability, ultimately enabling us to use the SWS variability to infer its underlying tissue microstructure. Different combinations of top three features were tested to understand the sensitivity of our parameter selection.<i>Main results</i>. Even though volume fraction and the SWS estimates were highly correlated, percent inclusion by itself as a single predictive factor was not an accurate indicator of the SWS estimates. For the sFLP tissue models developed for the current study, none of the individual SC features were able to predict the SWS estimates. Regardless of the top features identified, the model prediction correlation remained constant for each prediction iteration. However, even though the top three features across the five ML-based prediction iterations had different specific names, the features were all highly correlated.<i>Significance</i>. The findings from our current study suggest that while the percent inclusion rate was highly correlated to the mean SWS and SWS-STD, the percent inclusion rate alone cannot predict mean SWS or SWS-STD. mean SWS and SWS-STD provide unique information regarding the sFLP tissue microstructure, and both SWS estimates should be considered when analyzing fibrotic liver tissue. Consistent feature identification with previous published studies demonstrated that the sFLPs developed for this study may be representative of real-world patient data.