KEY POINTS
- This multicentre retrospective study included 1,077 patients with resected oral tongue squamous cell carcinoma: 658 in the derivation cohort and 419 in two external validation cohorts.
- A Vision Transformer model analysed preoperative T2-weighted magnetic resonance imaging. The best-performing input included the primary tumour and a 10 mm peritumoural extension.
- The integrated pathological deep-learning nomogram combined the imaging score with pT stage, pN stage, tumour grade and tumour budding.
- In the internal validation cohort, the integrated model achieved concordance indices of 0.809 for overall survival and 0.726 for disease-free survival, outperforming TNM staging.
- In the two external cohorts, concordance indices ranged from 0.778 to 0.804 for overall survival and from 0.736 to 0.737 for disease-free survival.
- In the favourable-risk group, postoperative radiotherapy was not significantly associated with overall survival (hazard ratio 0.429, p = 0.220) or disease-free survival (0.899, p = 0.764).
- In the poor-risk group, postoperative radiotherapy was associated with improved overall survival (hazard ratio 0.194) and disease-free survival (0.332), both p < 0.001. Treatment allocation was not randomized.
CLINICAL TAKEAWAY
Combining preoperative magnetic resonance imaging with pathological factors may improve selection for postoperative radiotherapy beyond conventional staging. However, the model cannot yet justify omitting radiotherapy in favourable-risk patients because the treatment comparisons remain vulnerable to selection bias and residual confounding.