MESAHA-Net: Multi-Encoder Self-Adaptive Hard Attention Network

Paper: Multi-Encoder Self-Adaptive Hard Attention Network with Maximum Intensity Projections for Lung Nodule Segmentation
Journal: Computers in Biology and Medicine, 2025
Authors: Muhammad Usman, Azka Rehman, Abd Ur Rehman, Abdullah Shahid, Tariq Mahmood Khan, Imran Razzak, Minyoung Chung, Yeong-Gil Shin

---

Overview

Accurate lung nodule segmentation is critical for early-stage lung cancer diagnosis.
MESAHA-Net introduces a multi-encoder, self-adaptive hard attention mechanism combined with bidirectional Maximum Intensity Projections (MIPs) to achieve state-of-the-art performance with lightweight computation.

• Triple Encoder: Raw slice + forward MIP + backward MIP
• Adaptive Hard Attention: ROI-guided, eliminating rescaling-induced errors
• Lightweight Design: ~0.44M parameters with ~48 ms per slice inference
• High Accuracy: Consistently outperforms Res-UNet, DEHA-Net, CMSF, and other baselines

---

Architecture

MESAHA-Net integrates contextual 2D and 3D information via multi-input encoders.
A self-adaptive hard attention block selectively amplifies nodule regions while suppressing irrelevant background features.
This leads to more precise 3D segmentation across heterogeneous lung nodules.

---

Results

LIDC-IDRI Dataset

• DSC: 88.27 ± 7.42
• Sensitivity: 92.88 ± 9.54
• PPV: 86.95 ± 11.29

LNDb Dataset

• DSC: 82.17
• Sensitivity: 85.96
• PPV: 86.57

Efficiency

• Parameters: ~0.44M
• Inference time: ~48 ms per slice

MESAHA-Net outperforms existing methods while being faster and lighter.

Key Advantages

✔ Eliminates rescaling-induced errors with adaptive hard attention
✔ Handles small and heterogeneous nodules better than prior models
✔ Provides robust generalization across datasets (LIDC-IDRI, LNDb)
✔ Clinically feasible: real-time inference on standard GPUs

---

Citation

If you use this work, please cite:

@article{MESAHA-Net-2025,
  title   = {Multi-Encoder Self-Adaptive Hard Attention Network with Maximum Intensity Projections for Lung Nodule Segmentation},
  author  = {Muhammad Usman and Azka Rehman and Abd Ur Rehman and Abdullah Shahid and Tariq Mahmood Khan and Imran Razzak and Minyoung Chung and Yeong-Gil Shin},
  journal = {Computers in Biology and Medicine},
  year    = {2025},
  note    = {Preprint}
}