Prediction of Tumor Budding Grading in Rectal Cancer Using a Multiparametric MRI Radiomics Combined with a 3D Vision Transformer Deep Learning Approach.

Rationale and objectives: The objective is to assess the effectiveness of a multiparametric MRI radiomics strategy combined with a 3D Vision Transformer (ViT) deep learning (DL) model in predicting tumor budding (TB) grading in individuals diagnosed with rectal cancer (RC).

Materials and methods: This retrospective study analyzed data from 349 patients diagnosed with rectal adenocarcinoma across two hospitals. A total of 267 patients from our institution were randomly allocated to a training cohort (n=187) or an internal test cohort (n=80) in a 7:3 ratio. Furthermore, a cohort of 82 patients from another hospital was established for external testing purposes. Univariate and multivariate analyses were performed to pinpoint independent clinical risk factors, which were then utilized to develop a clinical model. Radiomics (Rad) models, a 3D ViT DL model, and a combined model (DLR) were built using 3D T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and contrast-enhanced T1-weighted imaging (T1CE). The evaluation of each model's predictive performance involved calculating the area under the curve (AUC), conducting the Delong test, and examining calibration curves alongside decision curve analysis (DCA).

Results: No notable clinical characteristics were observed in either univariate or multivariate analyses, hindering the establishment of a clinical model. The DLR model demonstrated exceptional performance, attaining an AUC of 0.938 (95% CI: 0.906-0.969) within the training cohort, 0.867 (95% CI: 0.779-0.954) in the internal test cohort, and 0.824 (95% CI: 0.734-0.914) in the external test cohort.

Conclusion: The combination of multiparametric MRI radiomics and 3D ViT DL effectively and non-invasively predicts TB grading in RC patients, offering valuable insights for personalized treatment planning and prognosis assessment.