Auto-LSN: fully automated liver surface nodularity quantification in CT based on deep learning for the evaluation of advanced chronic liver disease.
Authors
Affiliations (13)
Affiliations (13)
- Guerbet Research, Villepinte, France. [email protected].
- MAP5 (UMR 8145), Université Paris Cité, Paris, France. [email protected].
- Radiology B Department, Hôpital Cochin, Paris, France. [email protected].
- Radiology Department, Hôpital Beaujon, Clichy, France.
- INSERM U1149 Centre de Recherche sur l'Inflammation (CRI), Université Paris Cité, Paris, France.
- Radiology Department, CHU Grenoble Alpes, La Tronche, France.
- Univ. Grenoble Alpes, Saint-Martin-d'Hères, France.
- Guerbet Research, Villepinte, France.
- Hepato-Gastroenterology and Digestive Oncology Department, CHU Grenoble Alpes, La Tronche, France.
- Institute for Advanced Biosciences, CNRS UMR 5309-INSERM U1209, Grenoble, France.
- MAP5 (UMR 8145), Université Paris Cité, Paris, France.
- Radiology Department, Centre Hospitalier Universitaire Angers, Angers, France.
- HIFIH Laboratory, EA 3859, Université d'Angers, Angers, France.
Abstract
Liver surface nodularity (LSN) is a recognized non-invasive biomarker of cirrhosis. This study introduces auto-LSN, an artificial intelligence (AI)-based algorithm for fully automated LSN quantification, assesses its association with fibrosis stage and its non-inferiority in diagnostic performance for advanced chronic liver disease (ACLD) and cirrhosis compared to the FDA-approved, semi-automated liver boundary analysis (LBA) software. This retrospective, bicentric study included patients with chronic liver disease risk factors who underwent CT and liver biopsy between April 2014 and March 2020. Fibrosis stages were grouped into F3-F4 (ACLD) vs F0-F2, and F4 (cirrhosis) vs F0-F3 per the METAVIR. LSN was measured with auto-LSN and LBA. Their association with fibrosis grade and diagnostic accuracy for ACLD and cirrhosis were compared using a -0.05 non-inferiority margin. Mann-Whitney-Wilcoxon tests, Spearman correlation, and area under the receiver operating characteristic curve (AUC) were used. In 127 patients (68 ± 12 years; 97 men), auto-LSN demonstrated a positive correlation with fibrosis stage (ρ = 0.59; 95% CI [0.48, 0.68]), similar to LBA (ρ = 0.44; 95% CI [0.32, 0.55]), both p < 0.001, with differences within the non-inferiority margin ([0.03, 0.26]). Auto-LSN achieved AUCs of 0.79 (95% CI [0.70, 0.87]) for ACLD and 0.84 (95% CI [0.76, 0.91]) for cirrhosis, comparable to LBA's AUCs of 0.73 (95% CI [0.64, 0.82]) and 0.74 (95% CI [0.66, 0.83]), respectively. All differences were within the non-inferiority margin. Auto-LSN correlates positively with fibrosis stage and provides non-inferior diagnostic performance compared to LBA. Its full automation and accuracy support its potential for opportunistic screening and objective patient monitoring. Question LSN is a key radiological feature for non-invasive ACLD diagnosis. However, current LSN quantification software is only semi-automated, thus time-consuming. Findings The fully automated auto-LSN algorithm for LSN quantification achieved statistically non-inferior diagnostic performance compared to existing semi-automated software for the detection of ACLD and cirrhosis. Clinical relevance Auto-LSN, as a fully automated solution, offers a reliable alternative to existing semi-automated software, enabling mass opportunistic screening of the general population-by evaluating all CT scans performed for any indication-and supporting objective follow-up of at-risk patients.