Computed Tomography-Derived Sarcopenia and Two- Versus Three-Dimensional Body Composition for Prognostication in Colorectal Cancer Patients Receiving Chemoradiotherapy: An Automated Deep Learning Segmentation Study with External Reproducibility Assessment

Abstract

Background: Skeletal muscle depletion predicts poor outcomes in gastrointestinal cancers, but whether volumetric three-dimensional (3D) body composition adds anything over the standard two-dimensional (2D) single-slice approach in colorectal cancer (CRC) patients receiving chemoradiotherapy is unclear. We quantified CT-derived sarcopenia and directly compared 2D and 3D body composition metrics from a fully automated deep learning pipeline. Methods: We retrospectively analyzed 368 patients with CRC. Body composition was quantified automatically from CT using SMAT-BC, a pipeline combining TotalSegmentator-based vertebral localization with an nnU-Net Residual Encoder XL network incorporating a Transformer bottleneck for four-class tissue segmentation. Single-slice L3 (2D) and L1–L5 volumetric (3D) indices were derived. The primary endpoint was overall survival (OS); recurrence-free survival (RFS) was secondary. Cox proportional hazards models with bootstrap optimism correction were used. Measurement reproducibility was assessed in 25 external TCGA-COAD cases. Results: Sarcopenia was strongly associated with both overall and recurrence-free survival (unadjusted OS HR 2.16, 95% CI 1.55–3.00; unadjusted RFS HR 1.83, 95% CI 1.36–2.47; both raw p < 0.001; adjusted OS HR 1.89, 95% CI 1.57–2.28; adjusted RFS HR 1.44, 95% CI 1.22–1.70; FDR-adjusted p < 0.0001 for both). L3 single-slice indices were strongly correlated with volumetric indices (r = 0.91 for muscle index) and added discriminant value over the clinical model for overall survival (optimism-corrected C-index: clinical 0.602 (95% CI 0.578–0.626), clinical + 2D 0.631 (95% CI 0.609–0.654), clinical + 3D 0.599 (95% CI 0.574–0.624); DeLong p = 0.002 for clinical + 2D vs. clinical, FDR-adjusted p = 0.006). The 2D- and 3D-augmented models yielded overlapping bootstrap confidence intervals and were not clinically meaningfully different in this cohort (OS ΔC = +0.032, 95% CI +0.013 to +0.051; FDR-adjusted p = 0.006; RFS ΔC = +0.008, 95% CI −0.011 to +0.027; FDR-adjusted p = 0.612). Conclusions: Automated CT-derived sarcopenia is an independent predictor of survival in CRC patients receiving chemoradiotherapy. In our cohort, single-slice L3 measurement matched or exceeded volumetric discrimination, but the 2D- and 3D-augmented models yielded overlapping bootstrap confidence intervals for both endpoints: for overall survival, the DeLong FDR-adjusted p value for the 2D-versus-3D contrast was 0.006, and 0.612 for recurrence-free survival. Because no non-inferiority margin was pre-specified, the 2D–3D comparison is presented as exploratory, and we make no formal claim of non-inferiority or equivalence for either endpoint. The findings support single-slice L3 measurement as an efficient biomarker for risk stratification but warrant external validation in larger prospectively designed cohorts.