Discussion

The noninvasive prognostic assessment of rectal cancer (RC) remains a significant clinical challenge [29, 30]. In particular, accurate preoperative identification of perineural invasion (PNI) is critical for personalized RC management, as PNI-positive patients often require intensified neoadjuvant chemoradiotherapy (nCRT) to improve survival [31–33]. We hypothesized that a framework combining super-resolution MRI, 2.5D deep learning, and peritumoral radiomics would outperform conventional methods. Our results validated this, with the combined model achieving superior discriminative performance. It is worth noting that the high AUC in the training set (0.993) is an expected outcome of the cascaded fusion architecture, where the final classifier learns from the outputs of sub-models already optimized on the training data. Importantly, the performance in the external testing cohort (AUC 0.868) suggests that the model exhibits promising generalizability and has avoided overfitting. This robustness is underpinned by our rigorous implementation of strictly isolated validation cohorts to prevent data leakage, complexity control via LASSO regularization, and data augmentation strategies (e.g., random cropping and flipping) during the deep learning training process.

This study investigated the utility of a 2.5D-GAN hybrid architecture. By utilizing orthogonal tomographic sections and super-resolution reconstruction, we mitigated the Z-axis resolution limitations of standard MRI [34]. ResNet101 was identified as the optimal backbone, balancing sensitivity and specificity. The subsequent multi-instance learning (MIL) elevated diagnostic precision, while gradient-weighted class activation mapping (Grad-CAM) elucidated region-specific saliency at tumor-neural interfaces, enhancing interpretability. Notably, tumor length was established as an independent PNI predictor. While statistically significant, the modest OR (1.062) of tumor length implies limited standalone clinical utility. However, it acts as a stable macroscopic baseline complementing microscopic imaging features; its integration into the Combined Model enhanced generalizability (AUC 0.868) over deep learning alone (AUC 0.812), thereby justifying its inclusion. These insights collectively underscore the potential of hybrid radiomic-deep learning paradigms.

Our methodology introduces three principal innovations. First, we addressed MRI resolution anisotropy via a 2.5D-GAN architecture. Regarding the super-resolution reconstruction, a primary concern in GAN-based methods is the potential introduction of artificial textures or “hallucinations.” However, our study provides robust evidence of feature stability: the extracted radiomic signature achieved high predictive accuracy in the independent external testing cohort (AUC = 0.868). This confirms that the super-resolution process enhanced genuine, generalizable tumor heterogeneity rather than generating random noise, thereby ensuring the reliability of the downstream analysis. Second, our 2.5D approach synergizes 2D efficiency with 3D fidelity. Zhang et al [19] demonstrated that 2.5D architectures require fewer parameters than 3D CNNs while preserving spatial context. Similarly, Chakrabarty et al [35] successfully employed 2.5D models for glioma genotyping. By aggregating orthogonal planes, our framework captures interslice correlations [36] without the prohibitive computational costs of 3D models [37]. A notable challenge in applying 2.5D deep learning to PNI prediction is the potential for label noise introduced by slice-level labeling. Because ground-truth PNI status is determined at the patient level via pathology, assigning this global label to every MRI slice implies that all cross-sections of a PNI-positive tumor exhibit predictive features. In reality, PNI is heterogeneous; specific slices may not visually depict neural invasion, creating a “weakly supervised” learning scenario. We addressed this by implementing a Multi-Instance Learning (MIL) strategy combined with Ensemble Fusion. By treating the patient as a “bag” of slices and aggregating features via prediction likelihood histograms (PLH) and Bag-of-Words (BoW) alongside ensemble pooling, the model focuses on the distributional patterns of malignancy across the entire tumor volume. This aggregation effectively filters out noise from non-informative slices, enabling robust patient-level predictions. Third, unlike prior studies focusing solely on intratumoral features [16, 17, 24–27], we integrated the extramural microenvironment. Since neural structures often localize beyond the muscularis mucosae, extramural PNI correlates with dismal survival [28]. We systematically analyzed the intratumoral and five concentric peritumoral zones. The 2 mm peritumoral model (Peri2mm) achieved optimal performance. Anatomically, this selection reflects a balance between sensitivity and specificity. A 2 mm margin captures the immediate “invasive front” and the tumor-host interface relevant to perineural invasion, whereas a narrower margin (1 mm) may miss subtle microenvironmental changes. Conversely, given the limited thickness of the mesorectum, wider margins (3–5 mm) risk including non-specific tissues—such as the mesorectal fascia or healthy adipose tissue—thereby introducing background noise that dilutes tumor-specific signals. Technically, the application of super-resolution GAN was pivotal for this analysis; by minimizing partial volume effects and sharpening tumor boundaries, the SR reconstruction ensured that this precise 2 mm zone could be delineated with high geometric fidelity, despite the potential concerns regarding generative artifacts. Comparatively, while Li et al [38] developed a CT-based nomogram, our MRI-based framework offers superior soft-tissue contrast without ionizing radiation. Furthermore, our 2.5D deep learning architecture overcomes the inherent limitations of traditional radiomics, such as poor interpretability and manual feature dependency [39]. Finally, multicenter validation across heterogeneous scanners enhanced model robustness, a critical advance over single-center studies.

Several limitations warrant consideration. First, this is a retrospective two-center study. While the multicenter cohort (N = 312) is relatively large, the external testing cohort (n = 66) is relatively small, which may limit the ability to definitively claim broad clinical generalizability. Second, the current protocol relies on manual ROI delineation, which may introduce interobserver variability and increase the workload in clinical practice. We aim to address this by integrating automated segmentation algorithms in future work. Third, while the model shows promising predictive performance, this study lacks a prospective clinical impact assessment. Therefore, the ability of the model to directly influence neoadjuvant therapy decisions and improve patient outcomes remains to be validated. Large-scale prospective clinical trials are essential to confirm its clinical utility and facilitate its translation into routine practice.

In conclusion, we established a noninvasive framework harmonizing 2.5D DL, super-resolution MRI, and dual-region radiomics for PNI prediction. The combined model demonstrates promising calibration and predictive generalizability, outperforming existing methods. Clinically, this tool may assist in the identification of high-risk patients, potentially supporting personalized neoadjuvant therapy decisions. However, further large-scale prospective validation is required to confirm its clinical impact.