Hybrid-View Attention Network (HVAN)

This repository contains the official implementation of the paper:

"Hybrid-View Attention Network for Clinically Significant Prostate Cancer Classification in Transrectal Ultrasound"
Accepted at MICCAI 2025 [[MICCAI2025]]

---

🧠 Method Overview

Overall Framework

The proposed HVAN framework introduces a hybrid-view learning strategy for transrectal ultrasound (TRUS) images. It consists of:

• CNN-Transformer Hybrid Backbone: Combines convolutional layers for local feature extraction with a transformer-based Hybrid-View Attention (HVA) module for modeling long-range dependencies. See panel (a).
• Hybrid-View Attention (HVA): Enhances representations through both intra-view attention (refining features within a view) and cross-view attention (capturing complementary information across views), as illustrated in panels (b) and (c).
• Hybrid-View Adaptive Fusion (HVAF): Dynamically integrates multi-scale features along spatial and channel dimensions for improved classification accuracy.

---

Hybrid-View Adaptive Fusion (HVAF)

---

📊 Results

Comparative and ablation results prove the efficacy of our method.

---

🚀 Training

To train the network, simply run:

python train.py