Anatomy-Privileged Distillation with Token Routing for MRI-Based Prediction of Perineural Invasion

2026-07-13

AI Summary

Preoperative non-invasive prediction of perineural invasion (PNI) in intrahepatic cholangiocarcinoma remains challenging, as existing approaches rely on manual annotations or contrast-enhanced imaging. This work proposes an anatomy-guided teacher–student distillation framework that, for the first time, integrates anatomical priors into a lightweight Vision Transformer via a token routing mechanism. The teacher model leverages tumor and liver masks to learn dense token routing, while the student model distills and aggregates critical tokens under a fixed computational budget, requiring no anatomical masks during inference. Using only T2-weighted MRI, the method achieves an average AUROC of 0.750 across 155 patients, with an inference time of 8.02 ms and a computational cost of merely 1.43 GFLOPs on a Jetson Orin Nano, demonstrating a favorable balance between performance and clinical practicality.

Abstract

Perineural invasion (PNI) is associated with poor postoperative outcomes in intrahepatic cholangiocarcinoma, but it is confirmed by surgical pathology. Existing preoperative imaging models often rely on radiologist-defined variables, contrast-enhanced imaging, or manual annotations. We propose an anatomy-privileged teacher--student framework for patient-level PNI prediction from T2-weighted MRI. During training, the teacher uses MRI with tumor and liver masks to learn dense token routing, and the student distills this guidance to retain and aggregate informative tokens under a fixed budget. Anatomical supervision is restricted to training, and the deployed model does not require masks at inference. In 155 patients, the proposed method achieved the highest mean AUROC of 0.750 among matched MRI-only baselines evaluated under the same protocol, with 1.43 GFLOPs and 8.02 ms per case on a Jetson Orin Nano Super Developer Kit.

Problem

• Perineural Invasion
• MRI
• Preoperative Prediction
• Intrahepatic Cholangiocarcinoma
• Anatomy-Privileged Learning

Innovation

• anatomy-privileged distillation
• token routing
• MRI-based PNI prediction
• mask-free inference
• efficient deep learning