Abstract
Abstract
Background
We evaluated whether CT intensity and texture radiomics provide predictive value beyond simple volume dynamics for lesion-level response prediction in metastatic melanoma treated with immune checkpoint inhibitors.
Methods
We conducted a retrospective multicenter study including 158 patients (1,626 lesions) from an internal cohort (
n
= 129 patients with standardized protocols) and an external cohort (
n
= 29 patients with heterogeneous protocols). The primary endpoint was binary lesion-level size response (clinical benefit vs. progressive disease) at the second follow-up. We trained and compared Random Forest classifiers to predict this endpoint using clinical variables, volumetric features, radiomics, and combined predictors derived from baseline and first follow-up imaging. Models were trained on the internal cohort using patient-grouped cross-validation and applied unchanged to the external cohort. Paired differences in AUC were assessed against a ± 0.05 equivalence margin.
Results
Baseline-only prediction of first follow-up response was near chance for all models. For predicting response at second follow-up, the volumetric model achieved an internal AUC of 0.84 [95% CI 0.78–0.88] and the radiomics model 0.86 [0.81–0.90]. In the external cohort, with 12 progression events, the volumetric and radiomics models achieved AUCs of 0.68 [0.50–0.85] and 0.68 [0.51–0.88], with no significant difference between any models. Organ-specific analysis showed no significant improvement from adding radiomics for any lesion site. Across all 288 radiomic features, the median absolute correlation with lesion volume was 0.37. Sensitivity analyses contained small or borderline gains that did not replicate externally.
Conclusions
Volumetric features capture the primary predictive information for early lesion-level response, and CT radiomics showed no conclusive improvement over volume dynamics. Confirmation in larger independent cohorts is required before radiomic signatures can be recommended for this task.
Clinical trial number
Not applicable.