1270 - Bridging the Gap In NPC PNI-Targeted Radiotherapy: AI-Driven Virtual Visualization of the Trigeminal Nerve Based on Anatomical Priors

Purpose/Objective(s): Precise target delineation is the cornerstone of head and neck cancer (HNC) radiotherapy, particularly given the paradigm shift toward adaptive "reduced-volume" strategies. However, microscopic perineural invasion (PNI)—often invisible on conventional imaging—constitutes a primary contraindication for volume reduction. This challenge is particularly critical in nasopharyngeal carcinoma (NPC). The trigeminal nerve (CN_V), the most common PNI pathway, exhibits discontinuous visibility on CT/MRI due to partial volume effects, necessitating subjective inference based on bony landmarks. We propose an AI-driven "virtual organ at risk; vOAR" framework utilizing anatomical priors to reconstruct the complete CN_V pathway, aiming to bridge the gap between anatomical inference and objective target definition.

Materials/Methods:

We constructed an nasopharyngeal carcinoma (NPC) dataset comprising 382 paired CT/MRI scans. A novel 5-class annotation protocol was developed to distinguish visible nerve segments, MRI-detectable nerve segments, invisible nerve segments, bony foramina (e.g., foramen ovale/rotundum), and adjacent structures. We designed a few-shot memory-augmented cascade framework featuring: 1. To overcome the challenges of ambiguous boundaries and limited training samples, we designed a memory-augmented cascade framework featuring a dual-stage architecture where CT-based landmarks provide rigid anatomical guidance. 2. A dynamic neural memory module incorporates few-shot learning strategies to store anatomical prototypes, enabling robust feature recognition in ambiguous regions despite limited training samples.

Results: Using 51 fully annotated paired cases, we performed 5-fold cross-validation. The framework demonstrated superior performance in delineating the radiologically challenging CN_V compared to baseline models.

Conclusion: This study validates the efficacy of integrating anatomical priors into deep learning for segmenting minute neural structures. By successfully "visualizing" radiologically invisible segments, this framework provides an objective tool for defining PNI risk volumes. This technology facilitates safer target volume reduction by ensuring occult PNI pathways are not inadvertently excluded and provides a geometric foundation for investigating dose-recurrence spatial relationships in high-risk neural regions.