A Deep Learning Approach for Hepatic Steatosis Detection and Grading from Standard B-mode Ultrasound Images
Tuesday, April 9, 2024
11:42am – 11:49am
Location: 412
Authors: Thodsawit Tiyarattanachai, Stanford University Krishna Bhatraju, Stanford University Jenny Vo-Phamhi, Stanford University Neha Antil, Stanford University Lindsey Marie Negrete, Stanford Aya Kamaya, Stanford Luyao Shen, Stanford University Ahmed Kaffas, Stanford
This presentation introduces a next-generation multi-instance deep learning (DL) strategy that effectively detects and grades liver steatosis in standard PACS-based B-mode ultrasound (US) images with multiple viewpoints and using magnetic resonance proton density fat fraction (MR-PDFF) as a reference.
Our work represents a significant step forward in DL for liver steatosis because it demonstrates the usefulness of a model built to optimize practicality for training and use in clinical settings; effectively the model can be run in the background and can be used to enhance radiology productivity by estimating steatosis grading automatically in standard of care abdominal ultrasound exams. Our model’s multi-class steatosis grading offers enhanced granularity compared to many prior models that typically use binary classifiers. Multiview point training improves the model's versatility and applicability in varied clinical settings. MR-PDFF is non-invasive, objective, and quantitative, and it offers a comprehensive representation of the entire liver. Our model's reliance on standard B-mode US images, which are standard-of-care in abdominal exams, makes the model easily deployable in various clinical settings; many other advanced models which employ MR-PDFF as a reference tend to use radiofrequency data, which is not as readily available.