EUS-FAT: Video-Based Multiple Instance Learning for Liver Fat Quantification from Endoscopic Ultrasound
Year: 2026
Type: conference-paper
Abstract: Accurate quantification of hepatic fat is critical for diagnosing non-alcoholic fatty liver disease (NAFLD). While magnetic resonance imaging-based proton density fat fraction (MRI-PDFF) is the current reference standard for non-invasive liver fat quantification, routine clinical adoption remains tempered by cost and limited availability. Endoscopic ultrasound (EUS), widely performed for other clinical indications, frequently captures liver tissue and thus offers a low-cost, readily available opportunity for opportunistic hepatic fat quantification. Our study systematically evaluates a video-based MIL framework with two complementary strategies: (a) a pretrained Swin-Tiny backbone with alternative attention mechanisms and (b) a universal ultrasound foundation model designed for cross-organ and cross-task image analysis. Within this framework, we propose a multi-head gated attention mechanism and a temporal-aware variation. Our results demonstrate that the Swin-Tiny + MIL framework substantially outperformed the foundation model + MIL framework. Swin-Tiny + MIL with the proposed multi-head gated attention achieved the best performance$(F 1=0.82, M A E=4.10)$, representing a novel contribution. In contrast, for the foundation model the best performance was achieved with the temporal convolution + multi-head gated attention module (F1 = 0.79, MAE = 4.79), highlight the importance of incorporating temporal information when relying on frozen embeddings. These findings suggest that MRI may not be required for PDFF quantification when EUS is already performed, enabling opportunistic screening, and highlight that tailored, task-specific architectures and attention mechanisms can outperform foundation models in specialized clinical applications. To the best of our knowledge, this represents the first application of MIL to EUS data for liver fat quantification
Institution: Mayo Clinic
Language: English
Subfield: Hepatology
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7172387962