Introduction

Rectal carcinoma (RC) remains a major global cancer burden, with rising incidence rates in recent decades [1, 2]. Despite diagnostic and therapeutic advances, many patients present with advanced disease and aggressive tumor biology that determines clinical outcomes [3]. Within prognostic determinants, perineural invasion (PNI) has gained prominence as a critical pathological marker. Defined histologically as malignant cell infiltration into neural structures—encompassing endoneurial invasion, > 33% circumferential nerve involvement, or perineurial penetration—PNI independently predicts postoperative recurrence, metastatic dissemination, and diminished survival [4, 5]. Its incorporation into AJCC TNM staging protocols reinforces its clinical significance, particularly in stratifying candidates for neoadjuvant chemoradiation (nCRT) and adjuvant chemotherapy [6, 7].

Current NCCN guidelines [8–10] maintain preoperative nCRT with total mesorectal excision (TME) as standard care for locally advanced RC, though survival outcomes remain suboptimal. Emerging European Society of Medical Oncology (ESMO) recommendations [11, 12] highlight PNI status as a crucial determinant for intensifying neoadjuvant/adjuvant regimens in high-risk Stage II patients. Paradoxically, while nCRT reduces PNI prevalence, postoperative histopathological assessment—the current diagnostic gold standard—fails to characterize pretreatment tumor biology, thereby limiting preoperative risk stratification [13–15]. Conventional MRI and biopsy lack sufficient resolution to detect subtle neural infiltration, necessitating reliance on postoperative histology and delaying personalized therapy [16, 17]. This diagnostic void underscores the urgent need for reliable noninvasive biomarkers to preoperatively assess PNI, potentially revolutionizing treatment personalization and prognostic accuracy.

Contemporary deep learning (DL) advancements have demonstrated notable superiority in medical image analysis [18]. However, conventional 2D/3D approaches present inherent limitations: 2D methods neglect multiplanar contextual information, while 3D architectures demand prohibitive computational resources and extensive annotated datasets [19]. Addressing this dichotomy, Zhang et al [20] pioneered a 2.5D medical imaging paradigm employing orthogonal lesion-centered planes as multichannel inputs. This hybrid framework preserves 3D spatial relationships while circumventing computational bottlenecks, effectively balancing interslice correlation preservation with resource efficiency [19]. Subsequent validation studies, including Takao et al [21], report 2.5D models achieving 88.7% sensitivity with statistically significant false-positive reduction (p < 0.001) and enhanced positive predictive values (58.9%). A growing body of research further underscores its potential in medical image segmentation [21–23]. Although radiomic approaches have been employed for PNI prediction [16, 17, 24–27], existing efforts predominantly focus on intratumoral features, neglecting the prognostic significance of extramural perineural invasion (ePNI)—a critical microenvironmental component associated with dismal survival outcomes compared to mural perineural invasion (mPNI) or PNI-negative cases (5-year DSS: 26.4% vs. 63.7% vs. 78.1%) [28].

Notably, the synergistic application of 2.5D deep learning and intratumoral-peritumoral radiomic analysis remains unexplored. This study develops a hybrid radiomic-DL model that integrates dual-region features with super-resolution imaging, investigating the clinical potential of a 2.5D DL-based framework for preoperative PNI prediction in RC.