Incremental prognostic value of CT radiomics and deep learning beyond complete tumor volume in non-small cell lung cancer: a reanalysis of LUNG1

The incremental value of CT radiomics and deep learning for overall survival prediction requires comparison with tumor volume measured from the complete gross tumor volume (GTV). Preprocessing and model fitting must remain confined to training patients. We reanalyzed 420 patients in NSCLC-Radiomics (LUNG1), of whom 372 died. We compared complete-GTV volume, 107 original-image radiomic features, a fixed pretrained 3D CNN representation and feature-level fusion on identical held-out five-fold assignments. The volume, clinical, radiomic, frozen-CNN and fusion models were evaluated over 20 repetitions of five-fold cross-validation. We assessed Harrell’s C-index, the integrated Brier score from 0.5 to 2 years, two-year calibration and decision curves. GTV coverage by a fixed 96-mm cubic crop was assessed in every patient. ComBat harmonization within training folds was examined in a 295-patient subset with known reconstruction kernels. We used 46 patients with 14 deaths from NSCLC Radiogenomics only to assess exploratory transportability. The fixed crop excluded part of the GTV in 64 of 420 LUNG1 patients. On the initial five-fold assignment, mean C was 0.585 for complete-GTV volume, 0.567 for volume plus radiomics, 0.571 for clinical plus radiomics, 0.514 for the frozen CNN and 0.569 for feature-level fusion. Paired conditional intervals for the radiomics-minus-volume and fusion-minus-volume differences crossed zero. Across 20 repetitions, mean C was 0.588 for volume, 0.577 for volume plus radiomics and 0.568 for fusion. On the initial assignment, the corresponding integrated Brier scores were 0.221, 0.225 and 0.236. ComBat showed no clear improvement over volume or unharmonized radiomics in the known-kernel subset. The small Stanford cohort was too limited for comparative validation. In this cohort, the evaluated complete-GTV radiomic and frozen-CNN fusion models did not show a stable held-out survival prediction gain over complete-GTV volume. These findings support full-contour measurement, leakage-free validation and independent replication in a population-matched cohort.