Искусственный интеллект выявил, что врачи пропускают жировую болезнь печени в трети случаев при чтении КТ

TOPLINE An AI tool found that hepatic steatosis frequently went undocumented on routine noncontrast CT scans, with radiologists reporting the finding in only one third of scans that met imaging criteria. The finding suggests that the absence of steatosis in a radiology report should not be interpreted as evidence of its absence. METHODOLOGY - Researchers conducted a retrospective cross-sectional analysis of 3646 abdominal noncontrast CT scans from 2710 adult patients (mean age, 51.7 years; 47.5% women; BMI, 27.9) obtained between 2012 and 2020 to compare AI-detected steatosis against what radiologists documented in their reports. - They used an automated deep learning algorithm for liver segmentation and attenuation assessment to analyze CT images and measure mean volumetric liver attenuation for the detection of hepatic steatosis. - Image-based steatosis was defined as mean liver attenuation below 40 Hounsfield units (HU), whereas report-based steatosis was identified from radiology reports using text-based searches for steatosis-related terms; a hepatologist manually reviewed a subset of reports to confirm the text search accurately captured what radiologists had documented. - To independently validate the automated assessment, a blinded radiologist separately re-reviewed a random subset of 113 scans for the presence of hepatic steatosis using established imaging criteria. TAKEAWAY - The AI tool identified imaging-defined hepatic steatosis in 480 (13.1%) scans, with these scans showing significantly lower mean liver attenuation than scans without imaging-defined hepatic steatosis (29.8 vs 53.5 HU; P < .001). - Among the scans showing imaging-defined hepatic steatosis, only 32.7% had steatosis documented in the corresponding radiology report; scans with steatosis documented in radiology reports had significantly lower liver attenuation than scans with unreported steatosis (24.6 vs 32.4 HU; P < .001), suggesting that milder steatosis was more likely to be overlooked. - In multivariable analysis, lower liver attenuation (odds ratio [OR], 0.94; P < .0001), lower attenuation variability (OR, 0.88; P < .0001), and a higher BMI (OR, 1.04; P = .01) independently predicted steatosis reporting. - During the validation check — manual re-review of 57 scans with attenuation below 40 HU — confirmed steatosis in 35 (61.4%); only 19 of those (54.3%) had been noted in the original report, while 16 (45.7%) were missed initially and caught only on this re-review. IN PRACTICE “In this cross-sectional retrospective analysis of abdominal CT scans for multiple indications, we show that the presence of hepatic steatosis is often not reported,” the authors wrote. “We demonstrate how an AI tool can be applied to large imaging datasets in epidemiologic studies and suggest that incorporating AI-based tools to provide hepatic attenuation values in CT scans may improve the radiologist workflow and lead to higher steatosis reporting rates.” SOURCE The study was led by David Yardeni, Liver and Energy Metabolism Section, Liver Diseases Branch, National Institute of Diabetes and Digestive and Kidney Diseases, National Institutes of Health, Bethesda, Maryland. It was published online in Clinical Gastroenterology and Hepatology. LIMITATIONS The analysis relied on liver attenuation alone to identify steatosis, and other radiologic markers of steatosis were not considered. Researchers used a conservative attenuation threshold of 40 HU, which may have missed milder steatosis and potentially overstated reporting rates. Automated attenuation measurements may have been affected by segmentation errors or large low-attenuation liver lesions, which could have affected liver attenuation estimates. DISCLOSURES This study was funded by the Intramural Research Program of the National Institute of Diabetes and Digestive and Kidney Diseases and the National Institutes of Health Clinical Center. One author disclosed receiving royalties for patents or software licenses and institutional support through a cooperative research and development agreement with industry. This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.