MRI–endoscopy AI improves pathological complete response prediction after rectal chemoradiotherapy

In external validation, combined MRI–endoscopy AI predicted rectal cancer pathological complete response with AUC 0.818, but precision was only 37.8%.

KEY POINTS

  • This retrospective, two-center Chinese study developed a multimodal deep learning model combining rectal MRI and endoscopic imaging to predict pathological complete response (pCR) and disease-free survival (DFS) after neoadjuvant chemoradiotherapy for locally advanced rectal cancer. A total of 288 patients were included, comprising 124 training, 54 internal-validation, and 110 external-validation cases. All patients received 50–54 Gy in 25–28 fractions with concurrent capecitabine or XELOX/CAPOX, followed by total mesorectal excision approximately 6–12 weeks after treatment.
  • The model integrated complementary imaging information rather than relying on one modality. Post-treatment multiparametric MRI, including T2-weighted, diffusion-weighted, and contrast-enhanced T1-weighted sequences, was processed using a 3D convolutional neural network, while endoscopic images were analyzed using a vision transformer. Features were combined through a graph neural network, adaptive weighting, and cross-attention to generate separate pCR probabilities and DFS risk scores. MRI and endoscopic regions of interest required manual annotation, with interobserver ICC values of 0.87 and 0.85, respectively.
  • Multimodal imaging improved discrimination of pathological complete response in external validation. The combined model achieved an AUC of 0.818 (95% CI 0.698–0.928), compared with 0.781 for endoscopy alone and 0.747 for MRI alone. Corresponding AUCs for the combined model were 0.876 in training and 0.848 in internal validation. The observed pCR rate was 25.3% (45/178) at the development center and 17.3% (19/110) at the external-validation center.
  • External validation revealed important limitations in positive predictive performance. The multimodal model achieved 74.6% accuracy, 73.7% sensitivity, 37.8% precision, and an F1 score of 0.500 for pCR prediction. Although discrimination exceeded either single-modality model, the low positive predictive value indicates that a substantial proportion of patients classified as likely complete responders still had residual pathological disease. This limitation is particularly important when considering treatment decisions that might omit surgery.
  • The model also improved AUC for predicting DFS events, but not all classification metrics. External-validation DFS AUC was 0.794 (95% CI 0.686–0.888), compared with 0.748 for endoscopy and 0.739 for MRI. However, external-validation accuracy was 0.664 with multimodal AI versus 0.692 and 0.707 for endoscopy and MRI, respectively, while precision was 0.380 versus 0.426 and 0.385. Thus, the higher AUC did not translate into uniformly better classification performance across all reported endpoints.
  • Model-derived DFS risk groups demonstrated statistically significant prognostic separation. Patients were stratified using cohort-specific median predicted risk scores, with thresholds of 0.346 in training, 0.384 in internal validation, and 0.354 in external validation. Kaplan–Meier analysis showed significantly shorter DFS in the high-risk groups (log-rank p<0.01). However, DFS was converted into a binary event/no-event outcome for predictive modeling, and the study did not establish a clinically validated time-specific risk threshold or demonstrate improved outcomes through model-guided treatment decisions.
  • Clinical translation remains limited by retrospective selection and the absence of organ-preservation validation. Every included patient underwent radical surgery, meaning that the model predicted pathological response against resection specimens rather than prospectively identifying patients who could safely enter a watch-and-wait pathway. Only patients with complete high-quality MRI and endoscopic datasets were included, and differences in age, nodal stage, imaging acquisition, and institutional practices may affect generalizability. Prospective validation, predefined decision thresholds, and comparison with multidisciplinary clinical complete response assessment are required.

CLINICAL TAKEAWAY

Combining MRI and endoscopic imaging with deep learning may improve preoperative pathological response assessment and prognostic stratification in locally advanced rectal cancer compared with either imaging modality alone. However, the modest positive predictive value and lack of prospective clinical validation mean that this model cannot currently support decisions to omit surgery or independently select patients for watch-and-wait management.

SOURCE

Radiation Oncology

Browse more research Suggest a correction