Reducing operator-dependent variability in liver shear wave elastography through an AI-guided region-of-interest placement pipeline

To develop and evaluate an artificial intelligence (AI) pipeline that automates optimal circular region-of-interest (ROI) placement and the surrounding elastogram box, provides probe-angle/contact guidance, and measures skin-to-capsule distance (SCD) to standardize liver SWE acquisition, aiming to reduce operator-dependent variability.

A two-stage pipeline combined SimCLR self-supervised pretraining on 17,441 unlabeled ultrasound images, followed by ENet fine-tuning using 495 annotated images for liver segmentation and 1564 images for poor probe contact detection. Segmentation outputs were combined with constraints on valid liver tissue pixels, axial depth, lateral boundaries, and regions of poor acoustic coupling to generate masks for optimal ROI placement. A distance transformation algorithm then identified optimal ROI coordinates within these masks. An inter-observer reader study assessed AI performance by comparing automated ROI placement, SCD measurements, and probe-angle correction guidance with annotations from expert attending radiologists.

18,182 US images from 654 liver elastography examinations in 609 patients (mean age, 52.6 ± 12.95 years; 55.4% male) were acquired from three ultrasound systems. The AI pipeline achieved liver segmentation performance with a Dice coefficient of 0.89, sensitivity of 0.95, specificity of 0.93, and an area under the receiver operating characteristic curve (AUC) of 0.94. The reader study demonstrated significantly higher AI-to-attending radiologist agreement (93.8%) compared to inter-attending agreement for ROI placement (81%). Cohen's kappa agreement between two attending radiologists was 0.25.

The proposed AI-driven pipeline significantly improved ROI placement consistency and reproducibility, highlighting its potential to enhance clinical reliability in liver fibrosis assessment.

Question Operator-dependent variability in region-of-interest (ROI) placement compromises measurement reproducibility in B-mode shear wave elastography. We developed an AI pipeline to place an optimal ROI. Findings The AI pipeline achieved liver segmentation performance of Dice 0.89 (AUC 0.94) and demonstrated higher AI-to-attending agreement (93.8%) than inter-attending agreement (81%) for ROI placement. Clinical relevance By standardizing ROI placement, elastogram box selection, and probe angle/contact guidance, the AI pipeline reduces operator-dependent variability in liver SWE acquisition, supporting more reproducible liver stiffness measurements that could improve fibrosis staging reliability.