Validation of a Body-Composition Segmentation Software on a Diverse Public CT Scan Cohort

Body Composition, Obesity, Sarcopenia

Body Composition, CT, Computed Tomography, Segmentation, Deep Learning, Artificial Intelligence, Skeletal Muscle Index, Sarcopenia, Validation

This study evaluates the standalone performance of Soma, a deep-learning software developed by Nucleo Research, Inc. for the automated segmentation of body-composition tissues (skeletal muscle, subcutaneous adipose tissue, visceral adipose tissue, and intramuscular adipose tissue) on whole-body computed tomography (CT) images. The aim is to confirm that Soma produces segmentations and tissue-area measurements that agree with a multi-rater expert reference standard, on a diverse cohort representative of demographic and clinical variation. A total of 200 CT scans are sampled by stratified design from a curated pool of 2,066 scans aggregated from six publicly available, de-identified imaging datasets (autoPET, AMOS, MSD Pancreas, CT-ORG, ENHANCE.PET, RATIC). Three board-certified radiologists independently annotate the reference standard at the L3 slice. Primary performance is assessed using the Dice similarity coefficient against the multi-rater reference, with predefined thresholds and BCa bootstrap confidence intervals, both in aggregate and within every demographic and clinical subgroup. Secondary endpoints include Bland-Altman analysis of tissue-area agreement, 95th-percentile Hausdorff distance, Pearson correlation of derived indices, and Cohen's kappa for sarcopenia classification using Skeletal Muscle Index (SMI). The study is fully retrospective on de-identified images, involves no patient contact, and has been determined exempt by Salus IRB (Salus Number 26328) under 45 CFR 46.104(d)(4).

Background. Body composition derived from cross-sectional imaging is increasingly used to assess sarcopenia, cachexia, and metabolic risk across oncology, surgical, and metabolic conditions. Manual segmentation at the L3 vertebra is the established reference but is time-consuming and rater-dependent. Soma is a deep-learning software pipeline (U-Net segmentation of skeletal muscle, subcutaneous adipose tissue, visceral adipose tissue, and intramuscular adipose tissue; EfficientNet-Lite0 + BiLSTM for automated L3 slice detection) developed by Nucleo Research, Inc. to provide reproducible, automated body-composition measurements from routine abdominal CT. Objective. To validate the standalone segmentation performance of Soma against an expert multi-rater reference standard, in aggregate and across predefined demographic and clinical subgroups. Design. Retrospective, multi-source observational validation on 200 de-identified abdominal CT scans, drawn by stratified sampling from a curated pool of 2,066 scans across six publicly available datasets (autoPET, AMOS, MSD Pancreas, CT-ORG, ENHANCE.PET, RATIC). Stratification ensures representation across: BMI category (4 levels), age band (4 levels), sex (2 levels), body region (abdomen-only vs. whole-body), and clinical context (oncologic vs. non-oncologic). Reference Standard. Three board-certified radiologists independently annotate the four tissue classes on every fifth axial slice (stride of 5) across the full scan depth. Inter-rater agreement is summarized prior to consolidation; consensus reference is derived per pre-specified consolidation rules. Index Test. Soma processes each scan blinded to ground truth. Outputs include per-tissue segmentation masks, tissue cross-sectional areas, and downstream indices including Skeletal Muscle Index (SMI = muscle area / height\^2). Primary Endpoint. Mean Dice similarity coefficient between Soma and the multi-rater reference computed across all annotated slices, with predefined performance thresholds: greater than or equal to 0.90 for muscle, subcutaneous adipose, and visceral adipose tissues; greater than or equal to 0.85 for intramuscular adipose tissue. Thresholds must be met both in aggregate AND within every demographic or clinical subgroup with at least 20 scans. 95% bias-corrected and accelerated (BCa) bootstrap confidence intervals are reported. Secondary Endpoints. (1) Bland-Altman analysis of agreement on tissue cross-sectional areas (bias and 95% limits of agreement). (2) 95th-percentile Hausdorff distance per tissue class. (3) Pearson correlation coefficient for tissue areas and SMI. (4) Cohen's kappa for sarcopenia classification (binary, by sex-specific SMI cutoffs) with prespecified threshold of greater than or equal to 0.80. Safety, Privacy, and Ethics. The study is fully retrospective on previously collected, publicly available, de-identified imaging data, with no patient contact and no intervention. There is no foreseeable risk to subjects. The protocol has been determined exempt by Salus IRB (Salus Number 26328) under 45 CFR 46.104(d)(4) on 04 May 2026.

No related papers found yet.

United States