Super-resolution MRI and 2.5D deep learning for intratumoral-peritumoral radiomics in preoperative prediction of rectal cancer perineural invasion.
Authors
Affiliations (3)
Affiliations (3)
- Department of Magnetic Resonance Imaging Diagnostic, The Second Affiliated Hospital of Harbin Medical University, Harbin, China.
- Medical Imaging Center, Zhuhai People's Hospital (The Affiliated Hospital of Beijing Institute of Technology, Zhuhai Clinical Medical College of Jinan University), Zhuhai, China.
- Department of Magnetic Resonance Imaging Diagnostic, The Second Affiliated Hospital of Harbin Medical University, Harbin, China. [email protected].
Abstract
Preoperative prediction of perineural invasion (PNI) in rectal cancer (RC) is challenging due to limited MRI resolution and the neglect of extramural microenvironmental features. We developed a noninvasive framework integrating super-resolution MRI, 2.5D deep learning (DL), and intratumoral-peritumoral radiomics for preoperative PNI prediction. A dual-center cohort of 312 RC patients with pathologically confirmed PNI status was analyzed. Preoperative MRI underwent 4× super-resolution reconstruction using a generative adversarial network (GAN) to enhance tissue definition. Radiomic features were extracted from the tumor and peritumoral regions (1-5 mm). For the 2.5D DL model, the largest tumor cross-section and adjacent axial slices served as multichannel inputs to a ResNet101 backbone using transfer learning. Slice-level features were aggregated to patient-level predictions via multi-instance learning (MIL). Feature selection employed univariate analysis, Pearson correlation, mRMR, and LASSO regression. Model performance was validated via 5-fold cross-validation and an external testing cohort. The combined model, integrating MIL features, intratumoral radiomics, the optimal (2-mm) peritumoral radiomics, and clinical variables, achieved an area under the curve (AUC) of 0.912 (95% CI: 0.850-0.974) on internal validation and 0.868 (95% CI: 0.783-0.952) on external testing. Gradient-weighted class activation mapping (Grad-CAM) highlighted the tumor-neural interface, and tumor length was an independent predictor of PNI (OR = 1.062, p = 0.032). We propose a hybrid radiomic-DL framework leveraging super-resolution MRI and 2.5D spatial context to enhance preoperative prediction of PNI in RC. The model shows potential for risk stratification to support personalized neoadjuvant therapy decisions, with external testing suggesting promising generalizability. Question How can super-resolution MRI and 2.5D deep learning address the limitations of conventional imaging in predicting preoperative rectal cancer perineural invasion? Findings The integrated framework achieved an external testing AUC of 0.868, demonstrating superior performance compared to standalone radiomics and deep learning models. Critical relevance statement This study demonstrates a noninvasive framework integrating super-resolution MRI and 2.5D deep learning with intratumoral-peritumoral radiomics, offering a novel approach for preoperative rectal cancer perineural invasion prediction and potential assistance in personalized neoadjuvant therapy decisions.