This repository provides the official PyTorch implementation of EdgeS², a lightweight edge-disentangled spiral scanning framework for predicting difficult laryngoscopy (DL) from 3D head-and-neck CT volumes.
EdgeS² is designed to achieve high predictive performance with extremely low computational cost, making it suitable for deployment in resource-constrained clinical environments.
Difficult laryngoscopy (DL) is a major challenge in anesthetic airway management and a significant source of perioperative risk. While 3D head-and-neck CT enables objective preoperative DL assessment, existing learning-based methods are often limited by insufficient global anatomical modeling, implicit boundary representation, and high model complexity.
We propose EdgeS², a lightweight edge-disentangled spiral scanning framework for DL prediction from 3D CT volumes. EdgeS² adopts a local-to-global design, where a learnable decomposition module reduces spatial redundancy while preserving structural information. Critical airway boundaries are encoded via shallow edge-aware features, and long-range anatomical dependencies across axial, sagittal, and coronal planes are captured using a 3D spiral scanning Mamba mechanism. A global triple cross-attention module integrates edge-aware, non-edge-aware, and semantic representations.
Experiments on an in-house dataset of 499 patients demonstrate that EdgeS² achieves state-of-the-art performance (AUC 0.908, F1-score 0.833) with only 0.422M trainable parameters.
-
Edge-Disentangled Representation Learning
Explicitly separates boundary-sensitive and non-boundary features. -
Learnable Decomposition Module
Reduces spatial redundancy while preserving anatomical structure. -
3D Spiral Scanning Mamba (S²Mamba)
Implicitly captures long-range 3D anatomical dependencies with linear computational complexity, avoiding explicit multi-view feature extraction and fusion. -
Triple Cross-Attention Integration
Jointly aggregates edge-aware, non-edge-aware, and deep semantic features.
├── model.py # EdgeS² network architecture
├── dataset.py # Dataset loading and preprocessing
├── train.py # Training and validation script
├── requirements.txt # Required Python dependencies
│
├── framework.png # Network architecture illustration
├── visualization.png # Saliency and interpretability visualization
├── analysis.png # Quantitative analysis results
│
├── 4666073.nii.gz # Example 3D CT volume
├── 4671339.nii.gz # Example 3D CT volume
-
framework.png
Overall architecture of the proposed EdgeS² framework. -
visualization.png
2D and 3D saliency maps highlighting anatomically meaningful airway boundaries. -
analysis.png
Performance analysis and experimental results. -
*.nii.gz
Example CT volumes provided for demonstration and testing purposes.
Please install the required dependencies using: pip install -r requirements.txt
It is strongly recommended to strictly follow the provided dependency versions, as deviations may lead to inconsistent results.
To train and evaluate the EdgeS² model, run:
python train.py