AI-based adiposity segmentation showed high accuracy (Dice score 0.97; 95% CI 0.96-0.98) and faster processing than manual methods, though extreme between-study heterogeneity limits reproducibility.
Meta-Analysis (n=12,815)
Does AI-based segmentation provide accurate and reproducible quantification of visceral, subcutaneous, and total fat in adult women with overweight and obesity?
While AI enables highly efficient and accurate adiposity segmentation at an individual level, extreme protocol heterogeneity currently limits the clinical generalizability and reproducibility of these metrics across studies.
Effect estimate: Dice score 0.97 (95% CI 0.96-0.98)
ABSTRACT Artificial intelligence (AI) enables automated, high‐throughput adiposity quantification, offering refined risk stratification for women with overweight and obesity. We systematically reviewed and meta‐analyzed studies evaluating AI‐based segmentation of visceral, subcutaneous, and total fat in adult women populations (BMI: 25–29.9 and ≥ 30 kg/m 2 ), searching MEDLINE, CENTRAL, Embase, Scopus, ScienceDirect, and Web of Science from inception to January 15, 2025, in accordance with PRISMA guidelines. Studies were included if they (1) focused on adult women (≥ 18 years) undergoing MRI, CT, or DXA‐based fat quantification; (2) employed automated or conventional radiomics segmentation; and (3) reported quantitative adiposity metrics (volumes or areas). Two reviewers independently extracted data, assessed quality via QUADAS‐2, and synthesized results using inverse‐variance and sample size–weighted random‐effects models. Of 13,098 records screened, 38 studies (12,815 women) met inclusion criteria. AI methods demonstrated high segmentation accuracy, exemplified by 3D‐Unet (Dice score: 0.97 95% CI: 0.96–0.98) and > 99% faster processing than manual methods (0.05 vs. 46.5 min/patient; p < 0.001). However, the primary finding was extreme between‐study heterogeneity, indicating limited reproducibility of adiposity measurements across studies. Substantial heterogeneity undermined pooled volumetric estimates: subcutaneous fat volume averaged 22,536 cm 3 ( I 2 = 99.9%), visceral fat 3615 cm 3 ( I 2 = 100%), and total fat 35,062 cm 3 ( I 2 = 99.8%), with prediction intervals spanning anatomically implausible ranges. This variability stemmed from inconsistent acquisition parameters, segmentation protocols, and modality‐specific differences, with DXA‐derived subcutaneous fat volumes 5.6‐fold higher than MRI‐derived estimates and CT attenuation thresholds spanning 244 HU (Hounsfield Units). Subgroup analyses indicated 72% of variance arose from technical rather than anthropometric factors. While AI enables highly efficient and accurate segmentation at an individual level, protocol heterogeneity and underrepresentation of diverse populations limit clinical generalizability. Standardized imaging protocols, harmonized analytic frameworks, and inclusive sampling are essential to translate AI's precision into clinically reliable adiposity metrics for women's health.
Taddese et al. (2026) conducted a meta-analysis in Overweight and obesity (n=12,815). AI-based adiposity segmentation vs. Manual methods was evaluated on Segmentation accuracy (Dice score) (Dice score 0.97, 95% CI 0.96-0.98). AI-based adiposity segmentation showed high accuracy (Dice score 0.97; 95% CI 0.96-0.98) and faster processing than manual methods, though extreme between-study heterogeneity limits reproducibility.