Predicting Controlled Attenuation Parameter and Elastographic Modulus from Handheld Ultrasound Device Data using Machine Learning

Background: The growing global health burden of steatotic liver disease (SLD) requires accessible and cost-effective diagnostic tools. Current non-invasive methods, such as controlled attenuation parameter (CAP) and elastography modulus (e-Mod) estimation, rely on expensive elastography equipment, limiting their accessibility in resource-constrained environments. This study builds on previous work by using a larger patient cohort and applying advanced machine learning techniques, including transformer-based models, to estimate CAP and e-Mod from raw radiofrequency signals (RFS) acquired by handheld ultrasound (HUS) devices.

Methods: From a cohort of n=554 patients, raw ultrasound images were extracted using HUS(Clarius HD3, C3 & L15) and CAP and e-Mod were estimated using transient elastography.

Convolutional neural networks (CNNs) and transformer models will be applied for e-Mod and CAP prediction. Each model will be trained using a 5-fold cross-validation on a train set and evaluated on a previously separated test set, with a 1:1 train-test split.

Expected results: Previous experiments on a cohort of n=77 with the same training setup show a mean(standard deviation) R2=0.5(0.07) for CAP-value prediction, and R2=-0.6(0.3) for e-Mod prediction for CNN models.

We hypothesise that neural networks will be capable of predicting previously mentioned variables, once trained on the larger cohort. Especially for e-Mod, which has been a challenge in previous studies as models tend to overfit on smaller training sets. This approach could significantly improve the diagnosis of fibrosis and make liver disease screening more accessible in resource-constrained environments.

Publication History

Article published online:
20 January 2025