Adaptive Routing for Efficient Diffusion Transformer-Based PNI Prediction

2026-07-13

AI Summary

Accurately predicting perineural invasion (PNI) in cholangiocarcinoma from preoperative MRI scans remains highly challenging due to subtle imaging features that extend beyond the tumor boundary and the difficulty of modeling long-range spatial dependencies with conventional methods. To address this, this work proposes a diffusion-based classification framework that employs a Vision Transformer as the denoising network and incorporates an adaptive routing strategy—dynamically adjusting cross-attention heads, spatial tokens, and MLP width—to maintain sensitivity to peritumoral micro-patterns while substantially reducing computational cost. The proposed method achieves an AUC of 0.731 for PNI prediction in cholangiocarcinoma with a computational complexity of only 257.57 GFLOPs.

Abstract

Perineural invasion (PNI) is a critical prognostic factor in cholangiocarcinoma. However, its preoperative prediction from magnetic resonance imaging (MRI) remains challenging due to subtle imaging features that extend beyond tumor boundaries into surrounding regions. Conventional convolutional neural networks are limited in capturing long-range spatial dependencies. Transformer-based architectures improve global modeling of volumetric MRI by aggregating spatially distributed contextual cues, yet capturing subtle and noise-sensitive patterns in peritumoral regions remains challenging. Diffusion-based classifiers offer an alternative formulation by leveraging denoising-based class scoring to better capture such subtle patterns. However, these approaches introduce substantial computational overhead due to the combination of transformer-based modeling and iterative denoising processes. To address these challenges, we formulate PNI prediction as a diffusion-based classification problem and implement the denoising network using a transformer-based representation. To improve computational efficiency, we introduce adaptive routing across attention heads, spatial tokens, and MLP width. Experimental results demonstrate that the proposed approach achieves an AUC of 0.731 with 257.57 GFLOPs.

Problem

• Perineural invasion
• Cholangiocarcinoma
• MRI
• Subtle imaging features
• Computational efficiency

Innovation

• diffusion-based classification
• adaptive routing
• Transformer
• perineural invasion prediction
• computational efficiency