KEY POINTS
- This retrospective single-centre study included 60 patients receiving neoadjuvant chemoradiotherapy for rectal cancer between 2008 and 2024: 50 good responders and 10 poor responders, giving a poor-response prevalence of 16.7%.
- Response was classified as complete or partial response versus stable or progressive disease. Pathological findings were preferred, but imaging, stage change and medical-record documentation were used when pathological information was unavailable.
- A total of 1,148 radiomic features were extracted from the original clinical target volume on non-contrast planning computed tomography. The volume included the tumour-bearing rectum and treatment-relevant mesorectal, presacral, internal iliac and obturator regions rather than the gross tumour alone.
- Clinical variables included age, sex, clinical T and N stage, log-transformed carcinoembryonic antigen and tumour distance from the anal verge. Eighty percent received long-course treatment, median radiotherapy dose was 50.4 Gy, and 80% subsequently underwent surgery.
- Model development used ten repetitions of nested five-fold cross-validation, with scanner-based ComBat harmonization, feature selection and hyperparameter tuning confined to training folds. The clinical-only model achieved an AUROC of 0.507 and an area under the precision-recall curve of 0.155.
- The best combined model, Extra Trees, achieved an AUROC of 0.754 (95% confidence interval, 0.713–0.789) and an area under the precision-recall curve of 0.299. The combined LightGBM model achieved an AUROC of 0.740 and an area under the precision-recall curve of 0.313.
- At a post hoc rule-out-oriented threshold, the combined Extra Trees model achieved 94.0% sensitivity, 46.8% specificity, 26.1% positive predictive value and 97.5% negative predictive value. Its Brier score was 0.123 versus 0.264 for the clinical-only model, and decision-curve analysis suggested internal net benefit at low-to-moderate thresholds.
CLINICAL TAKEAWAY
Radiomics from an existing planning contour may provide a low-burden research pathway for pretreatment response stratification without requiring a separately segmented gross tumour. The apparent high negative predictive value should not guide treatment selection: only ten patients had poor response, thresholds were chosen post hoc, response assessment was heterogeneous and all performance estimates were internal.