CT Body Composition Radiomics Improves NSCLC Survival Prediction

Key Takeaways

• Among patients undergoing curative-intent NSCLC resection at three Chinese tertiary centers, the clinical–tumor–body composition model discriminated overall survival better than the clinical–tumor model in internal and independent external validation.
• Tumor-plus-body-composition Cox radiomics discriminated survival significantly better than tumor-only radiomics in both validation cohorts.
• Score-derived phenotypes differed in survival, with the body-composition-harmful phenotype having the poorest observed prognosis.

Survival after curative-intent resection for non–small cell lung cancer (NSCLC) can differ among patients with seemingly similar tumor characteristics. Presurgical computed tomography (CT) depicts not only the tumor but also skeletal muscle and adipose tissue, raising the question of whether host body composition adds prognostic information.

In a multicenter radiomics study, investigators retrospectively screened consecutive curative-intent resections for histologically confirmed NSCLC at three Chinese tertiary centers from January 2013 through December 2017. Of 2,880 screened patients, 1,038 met eligibility criteria, including contrast-enhanced chest CT within 30 days before surgery and complete clinical and follow-up data. The authors list a history of other malignancies among the exclusion criteria, although they also report between-center differences in a baseline variable termed 'history of malignancy.'

Centers A and B supplied a randomized 7:3 training and internal-validation split; Center C was held out for independent external validation. Overall survival ran from surgery to death from any cause or last contact in October 2023; median follow-up was 3.31 years. Investigators manually delineated tumors on venous-phase CT. Automated localization of the first through 12th thoracic vertebrae guided three-dimensional segmentation of skeletal muscle, subcutaneous fat, visceral fat, intermuscular fat, and bone. Tumor and body-composition radiomic risk scores were derived using screening, correlation filtering, and least absolute shrinkage and selection operator (LASSO) Cox selection. Tumor-only and tumor-plus-body-composition Cox radiomics models were assessed separately from clinical–tumor and clinical–tumor–body composition models built with extreme gradient boosting (XGBoost).

The integrated clinical–tumor–body composition (CTB) XGBoost model had a concordance index of 0.780 versus 0.725 for the clinical–tumor model in internal validation. At the independent center, the corresponding indices were 0.779 versus 0.726. Its time-dependent area under the curve (AUC) exceeded 0.80 at each assessed 1-, 2-, 3-, and 5-year survival horizon. AUC gains over the clinical–tumor model were significant at every assessed internal-validation horizon.

In the separate Cox radiomics comparison, tumor-plus-body-composition and tumor-only models had concordance indices of 0.749 versus 0.695 internally and 0.742 versus 0.692 externally; both differences were reported as significant (p < 0.05). In training, tumor and body-composition scores were independently associated with overall survival. Automated vertebral localization and tissue segmentation showed high agreement, although intermuscular fat segmentation lagged other compartments. Shapley additive explanations (SHAP) identified both scores as influential. Four score-derived phenotypes differed in survival across the three cohorts, with tumor-conservative patients showing the best observed prognosis. Decision-curve analysis favored CTB over clinical and clinical–tumor models for modeled net benefit across cohorts, not measured benefit from treatment decisions.

The retrospective design leaves room for selection bias, and overall survival is subject to residual confounding from adjuvant therapy, comorbidities, and molecular tumor factors. Baseline characteristics differed across centers, and additional independent external cohorts would clarify generalizability. In the internal-validation cohort, survival separation did not reach significance among stage III patients with N0–1 disease (p = 0.25). Tumor delineation remained manual despite automated tissue segmentation.

Huang and colleagues concluded that adding volumetric body-composition imaging features improved overall-survival discrimination beyond tumor features in their surgical NSCLC cohorts, including validation at an independent center.

Clinician Questions

Which patients were excluded from the NSCLC resection radiomics cohort?

For their cohort of patients undergoing curative-intent resection for histologically confirmed NSCLC, Huang and colleagues excluded those with a history of another malignancy or inadequate presurgical chest CT image quality or slice thickness.

How were harmful and conservative CT radiomics phenotypes assigned after NSCLC resection?

For patients with resected NSCLC, investigators compared the absolute SHAP contributions of the tumor and body-composition risk scores. The larger absolute contribution determined which score dominated; its positive or negative sign assigned harmful or conservative status, respectively.