Background Domain shift constrains the cross-site generalizability of radiomics-based survival prediction models. This study evaluated the transferability of five PET/CT radiomic feature groups in non-small cell lung cancer (NSCLC), examining three factors that may affect transferability: source-domain prognostic utility, distributional stability, and site-dependent prognostic association. We also applied an exploratory Domain-Divergence-Adjusted Utility (DDAU) score to assess whether feature-group transferability could be prioritized without target labels. Methods Bidirectional cross-site validation employed the TCIA NSCLC-Radiogenomics cohort from two institutions (VA, <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(\:n=88\)</EquationSource></InlineEquation>; Stanford, <InlineEquation ID="IEq2"><EquationSource Format="TEX">\(\:n=60\)</EquationSource></InlineEquation>). For each radiomic feature group (Shape, SUV, HU, Texture, and Wavelet), a least absolute shrinkage and selection operator (LASSO)-penalized Cox rad-score was derived in the source domain, added to a clinical baseline model (ClinX), and evaluated in the target domain. Performance was assessed using the C-index and integrated Brier score. Confidence intervals and p-values for C-index improvement were estimated using 1,000 paired bootstrap resamples. DDAU integrated source-domain C-index gain per feature with Kolmogorov–Smirnov (KS)-based distributional stability. Post-hoc concept-shift analysis used likelihood ratio tests for site × rad_score interactions. Results The Shape group showed positive numerical C-index improvement in both transfer directions (Stanford→VA: ClinX 0.656, ClinX+Shape 0.711, <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(\:\varDelta\:C=+0.056\)</EquationSource></InlineEquation>, 95% CI: −0.004 to + 0.120; VA→Stanford: ClinX 0.684, ClinX+Shape 0.740, <InlineEquation ID="IEq4"><EquationSource Format="TEX">\(\:\varDelta\:C=+0.055\)</EquationSource></InlineEquation>, 95% CI: −0.017 to + 0.130). DDAU showed positive descriptive rank tendencies with observed cross-site C-index in both directions (Spearman <InlineEquation ID="IEq5"><EquationSource Format="TEX">\(\:\rho\:=0.70\)</EquationSource></InlineEquation> for Stanford→VA and <InlineEquation ID="IEq6"><EquationSource Format="TEX">\(\:\rho\:=0.80\)</EquationSource></InlineEquation> for VA→Stanford). Secondary feature combinations and exploratory domain-adaptation comparisons did not show a consistent bidirectional advantage over the parsimonious ClinX+Shape model. Conclusions In this exploratory bidirectional PET/CT radiomics study, the Shape group showed the most consistent numerical C-index improvement across five feature groups evaluated for cross-site NSCLC survival prediction. DDAU is presented as a hypothesis-generating, target-label-free framework that makes source-domain utility and domain stability explicit and auditable before target-label evaluation; it is not claimed to outperform simpler utility-based ranking. Larger multi-institutional external validation is warranted.

Cross-site transferability of PET/CT radiomic feature groups for survival prediction in non-small cell lung cancer: a bidirectional analysis

期刊论文
出版模式：

开放获取
发表日期：

2026年8月5日

查看全文

Won Gi Choi,
Sun-Ho Kim,
Min Soo Kim

摘要

Background

Domain shift constrains the cross-site generalizability of radiomics-based survival prediction models. This study evaluated the transferability of five PET/CT radiomic feature groups in non-small cell lung cancer (NSCLC), examining three factors that may affect transferability: source-domain prognostic utility, distributional stability, and site-dependent prognostic association. We also applied an exploratory Domain-Divergence-Adjusted Utility (DDAU) score to assess whether feature-group transferability could be prioritized without target labels.

Methods

Bidirectional cross-site validation employed the TCIA NSCLC-Radiogenomics cohort from two institutions (VA, \(\:n=88\); Stanford, \(\:n=60\)). For each radiomic feature group (Shape, SUV, HU, Texture, and Wavelet), a least absolute shrinkage and selection operator (LASSO)-penalized Cox rad-score was derived in the source domain, added to a clinical baseline model (ClinX), and evaluated in the target domain. Performance was assessed using the C-index and integrated Brier score. Confidence intervals and p-values for C-index improvement were estimated using 1,000 paired bootstrap resamples. DDAU integrated source-domain C-index gain per feature with Kolmogorov–Smirnov (KS)-based distributional stability. Post-hoc concept-shift analysis used likelihood ratio tests for site × rad_score interactions.

Results

The Shape group showed positive numerical C-index improvement in both transfer directions (Stanford→VA: ClinX 0.656, ClinX+Shape 0.711, \(\:\varDelta\:C=+0.056\), 95% CI: −0.004 to + 0.120; VA→Stanford: ClinX 0.684, ClinX+Shape 0.740, \(\:\varDelta\:C=+0.055\), 95% CI: −0.017 to + 0.130). DDAU showed positive descriptive rank tendencies with observed cross-site C-index in both directions (Spearman \(\:\rho\:=0.70\) for Stanford→VA and \(\:\rho\:=0.80\) for VA→Stanford). Secondary feature combinations and exploratory domain-adaptation comparisons did not show a consistent bidirectional advantage over the parsimonious ClinX+Shape model.

Conclusions

In this exploratory bidirectional PET/CT radiomics study, the Shape group showed the most consistent numerical C-index improvement across five feature groups evaluated for cross-site NSCLC survival prediction. DDAU is presented as a hypothesis-generating, target-label-free framework that makes source-domain utility and domain stability explicit and auditable before target-label evaluation; it is not claimed to outperform simpler utility-based ranking. Larger multi-institutional external validation is warranted.