Multiparametric MRI-based multi-channel deep learning model for accurate preoperative prediction of perineural
Dalin Cheng1, Jingjing Zhao1, Zhanhong Liu1
1Lvliang First People's Hospital, Lvliang, China.
Objective:
To develop a multiparametric MRI-based multi-channel deep learning model for accurate preoperative prediction of perineural invasion (PNI) in patients with lymph node-negative rectal cancer (RC).
Methods:
This multicenter diagnostic study retrospectively enrolled 266 patients with RC from two institutions. Pretreatment multiparametric MRI sequences (T2WI, DWI, and CE-T1WI) were preprocessed, and for each sequence, the ROI was cropped into four views (roi-only, roi-all, enlarge-roi, and no-crop), which were then stacked into a 12-channel input for deep learning. Fusion across the four image sets generated five deep learning (DL) models, along with one hybrid radiomics-deep learning (DLR) model. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC, with stability assessed by stratified bootstrap resampling of 1,000 iterations), sensitivity, specificity, calibration curves, and decision curve analysis (DCA). The DeLong test with Bonferroni correction was used for pairwise model comparisons.
Results:
The dataset was divided into a training cohort (n = 134, Institution 1), an internal validation cohort (n = 58, Institution 1), and an external test cohort (n = 74, Institution 2). In the external test cohort, the conventional radiomics model achieved an AUC of 0.677 (95% CI: 0.412-0.941). The best-performing deep learning image model achieved an AUC of 0.843 (95% CI: 0.680-1.000). The combined DLR model further improved performance, demonstrating strong predictive ability, with AUCs of 0.928 (95% CI: 0.862-0.994) in the training cohort, 0.896 (95% CI: 0.797-0.995) in the internal validation cohort, and 0.863 (95% CI: 0.695-1.000) in the external test cohort. Bootstrap resampling (1,000 iterations) confirmed the stability of the DLR model, yielding consistent mean AUCs across all cohorts.
Conclusions:
This novel approach integrating radiomics features extracted from multiparametric MRI with multi-channel deep learning enables accurate preoperative prediction of PNI in RC, providing valuable support for individualized treatment planning and prognostic evaluation.