Table 2.
Performance of the local (LC) and loco-regional (LRC) control prediction models depending on the input. Combination of PT and LN radiomics improved prediction of LRC (p-value < 0.05). LC rad score: prediction of LC based on the PT radiomics.
| model input | endpoint | features | c-index 5-fold CV training | c-index validation (95% CI) |
|---|---|---|---|---|
| PT radiomics | LC | GLSZM zone entropy LLL NGTDM complexity LLL GLCM entropy | 0.81 | 0.70 (0.68–0.71) |
| PT radiomics | LRC | NGTDM complexity LLL NGTDM complexity LLL GLCM entropy | 0.67 | 0.63 (0.62–0.64) |
| LN radiomics | LRC | thickness SD spherical disproportion major axis histogram kurtosis | 0.72 | 0.60 (0.58–0.61) |
| PT + LN radiomics | LRC | LC rad score thickness SD spherical disproportion major axis histogram kurtosis | 0.75 | 0.67 (0.66–0.68) |
| PT radiomics distribution LNPT | LRC | Distribution LNPT feature not significant in the multivariate model | — | — |
| PT radiomics N stage | LRC | N stage not significant in the multivariate model | — | — |