From Anatomy to Pathology: How Well Can AI Interpret Static Point-of-Care Ultrasound Images?
Authors
Affiliations (3)
Affiliations (3)
- Department of Emergency Medicine, University of Health Sciences, Kanuni Sultan Suleyman Training and Research Hospital, Atakent Mahallesi, Turgut Ozal Boulevard 46/1, Kucukcekmece, Istanbul 34303, Türkiye.
- Department of Emergency Medicine, School of Medicine, Istanbul Medipol University, Kavacik Mahallesi, Ekinciler Caddesi No. 19, Beykoz, Istanbul 34810, Türkiye.
- Department of Emergency Medicine, University of Health Sciences, Haseki Training and Research Hospital, Ugur Mumcu Mahallesi, Belediye Street 7, Sultangazi, Istanbul 34265, Türkiye.
Abstract
<b>Background/Objectives</b>: Multimodal artificial intelligence may support point-of-care ultrasound (POCUS) interpretation in emergency medicine, but its performance on static images and sensitivity to acquisition context remain uncertain. This study evaluated GPT-5.6 Sol and assessed whether probe-region information improves interpretation across view/plane, anatomy, and pathology domains. <b>Methods</b>: In this paired comparative in silico study, 92 educational ultrasound images from seven anatomical regions were assessed under standardized uninformed (image only) and informed (image plus probe-region code) prompting conditions. Two independent raters scored responses against textbook-based references in three domains: view/plane identification, anatomical structure identification, and pathological finding detection. Each domain was scored from 0 to 2. Disagreements were resolved by a third ultrasound expert, and fully correct responses were defined as domain scores of 2. The primary outcome was the total performance score (0-6). Secondary outcomes included domain-specific scores, fully correct response rates, and exploratory subgroup analyses. Wilcoxon signed-rank and McNemar tests assessed paired scores and fully correct rates, respectively. <b>Results</b>: Total performance was higher with informed prompting (median (interquartile range), 5 (4-6) versus 4 (2-5); <i>p</i> < 0.001). Interrater agreement was substantial to almost perfect across evaluation domains (quadratic-weighted κ = 0.750-1.000). Fully correct rates increased from 58.7% to 88.0% for view/plane identification and from 47.8% to 82.6% for anatomical structure identification (both <i>p</i> < 0.001). Pathological finding detection improved from 35.9% to 48.9% (<i>p</i> = 0.002). Among pathology-positive images, missed or incorrectly labeled pathology remained frequent (44.8%) despite informed prompting. Regional differences in total performance were significant only under uninformed prompting (<i>p</i> = 0.002). <b>Conclusions</b>: Probe-region information substantially improved view/plane and anatomical structure identification, but pathological finding detection remained comparatively weak. In static POCUS interpretation, improved anatomical recognition did not translate into reliable pathology detection. Persistent false-negative and false-positive errors do not support independent diagnostic use. Potential applications in POCUS education and anatomical labeling require further validation.