Truace
Subscribe

Integrating Plain 2D Ultrasound and Clinical Data for Identification of At-Risk MASH: A Multicenter Proof of Concept Study

Background Identifying people with at-risk metabolic dysfunction-associated steatohepatitis (MASH) is a major challenge. We aimed to use deep learning to develop a convolutional neural network (CNN) model to facilitate the detection of at-risk MASH. Methods A multi-center cohort study was conducted across five tertiary hospitals in Japan and China. Clinical data and conventional ultrasound images were retrospectively collected. Organ segmentation using YOLO version 11 with quality control was performed, followed…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1948· cite & verify
Integrating Plain 2D Ultrasound and Clinical Data for Identification of At-Risk MASH: A Multicenter Proof of Concept Study
CT of cholecystitis by Mikael Häggström , M.D. Author info - Reusing images - Conflicts of interest: None Mikael Häggström , M.D. Consent note : Written informed consent was obtained from the individual, including online publication.. CC0 · http://creativecommons.org/publicdomain/zero/1.0/deed.en

In brief

Researchers built a deep learning pipeline that segments organs on plain 2D ultrasound images, then feeds them to an EfficientNet-B6 regression model. It predicts two composite risk scores, FAST and AGILE 3+, used to flag people with at-risk metabolic dysfunction-associated steatohepatitis. Data came from five tertiary hospitals in Japan and China.

The work matters because identifying at-risk MASH is difficult and transient elastography is not available everywhere. Routine ultrasound is more widely accessible, so a model that approximates elastography-based scores could widen detection. The study is labeled a proof of concept, so clinical benefit in practice has not been shown.

Main points

  1. The dataset had 416 training patients, 139 for internal validation, and 154 for external testing.
  2. Agreement with FAST was good (R2 0.684, category agreement 0.71); agreement with AGILE 3+ was moderate (R2 0.445, category agreement 0.63).
  3. FAST AUROC was 0.974 at the 0.35 cut-off and 0.932 at the 0.67 cut-off.
  4. AGILE 3+ AUROC was 0.908 at 0.451 and 0.902 at 0.679.

The gain

A CNN trained on conventional ultrasound images and clinical data estimated FAST and AGILE 3+ risk scores with high discrimination at rule-out and rule-in cut-offs, which may help where transient elastography is scarce.

The rundown

The pipeline used YOLO version 11 for organ segmentation with quality control, followed by an EfficientNet-B6 CNN regression model. Image-level predictions were averaged for each patient to estimate the risk categories.

Performance was judged on agreement across clinically relevant rule-out and rule-in thresholds, R2, root mean squared error, and AUROC. RMSE was 0.131 for FAST and 0.1964 for AGILE 3+.

Clinical data and ultrasound images were collected retrospectively across the participating hospitals. The authors conclude the approach may support detection in settings with limited access to transient elastography.

Sources

  1. Peer-reviewedJournal of Gastroenterology and Hepatology2026-10-16

The debate