MMA-Former: Multi-Window Mixture-of-Head Attention Transformer for Adaptive PNI Prediction in 3D MRI

2026-07-12

AI Summary

This study addresses the challenge of non-invasively predicting perineural invasion (PNI) in cholangiocarcinoma from 3D MRI, which requires simultaneous modeling of fine-grained details and global contextual information. To this end, the authors propose MMA-Former, an end-to-end 3D Transformer architecture featuring a parallel coarse-fine dual-path design, a multi-window partitioning strategy, and Window-Specific Mixture-of-Heads (WS-MoH) attention. The WS-MoH mechanism dynamically routes 3D windows to specialized or shared attention heads, enhancing local representation capacity and computational efficiency without increasing model parameters. Evaluated on a retrospective dataset of 168 T1-weighted MRI scans, MMA-Former achieves an AUC of 0.752, significantly outperforming state-of-the-art 3D CNN (AUC: 0.708) and Transformer baselines (AUC: 0.681).

Abstract

Perineural invasion (PNI) is a critical prognostic factor in cholangiocarcinoma. Non-invasive prediction from 3D MRI is challenging, demanding models that efficiently capture both fine-grained details and global context. We propose the Multi-window Mixture-of-Head Attention Transformer (MMA-Former), a novel end-to-end 3D architecture featuring a Coarse-Fine Transformer (CFT) structure for parallel multi-scale feature extraction. We advance this structure by integrating a novel Window-Specific Mixture-of-Head attention (WS-MoH) mechanism. Unlike standard Multi-Head Self Attention (MSA), WS-MoH generates a representation for each 3D window and dynamically routes the entire window to specialized or common attention heads. This enables spatially adaptive feature extraction tailored to the local context of each window, enhancing specialization and reducing redundancy without increasing parameters. Evaluated on a retrospective dataset of 168 T1-weighted MRI scans, MMA-Former achieved an AUC of 0.752, outperforming other 3D architectures, including the best CNN (AUC of 0.708) and Transformer baselines (AUC of 0.681).

Problem

• Perineural invasion
• 3D MRI
• Cholangiocarcinoma
• Non-invasive prediction
• Adaptive feature extraction

Innovation

• MMA-Former
• Window-Specific Mixture-of-Head attention
• Coarse-Fine Transformer
• adaptive feature extraction
• 3D MRI