📘 Paper (npj Digital Medicine, under revision)

Radiolucent foreign body aspiration (FBA) is challenging to diagnose due to its subtle imaging characteristics on CT. We present a deep learning pipeline that integrates 3D airway segmentation with multi-view classification. Our model was validated across three independent hospital datasets and outperformed expert radiologists in key diagnostic metrics.

• Segmentation: MedpSeg, a high-precision 3D U-Net-based segmentation model.
• Classification: ResNet-18 applied to 12-view 2D projections of the airway tree.
• Validation: Multi-center datasets from three hospitals in Wuhan, China.
• Comparison: Model demonstrated higher recall and F1 score than radiologists.

Requires Python 3.8+, PyTorch 1.10+

git clone https://github.com/ZheChen1999/FBA_DL.git
cd FBA_DL
pip install -r requirements.txt

• Airway segmentation: 3D U-Net (MedpSeg) with expert-in-the-loop refinement
• Snapshot generation: Multi-angle 2D views of 3D segmented airway
• Classification: ResNet-18 with Focal Loss and data augmentation
• Training strategy: 5-fold cross-validation, early stopping
• Hardware: Dual RTX A5000 GPUs for segmentation

1. Initial CXR (to rule out radiopaque FBA)
2. Chest CT if radiolucent FBA suspected
3. AI-assisted segmentation and classification
4. Use model results to prioritize bronchoscopy

@article{liu2025fba,
  title={Automated Detection of Radiolucent Foreign Body Aspiration on Chest CT Using Deep Learning},
  author={Liu, Xiaofan and Chen, Zhe and Tang, Zhiyong and Yang, Xun and Jiang, Yan and et al.},
  journal={npj Digital Medicine},
  year={2025},
  note={Under review}
}

Detailed methodology, datasets, training protocols, and ablation studies are provided in the Supplementary Materials. Topics include:

• Data Sources: Multi-center cohorts from Wuhan hospitals + ATM22 dataset
• MedpSeg Architecture: 3D U-Net with expert-guided iterative training
• ResNet-18 Classifier: Multi-view CNN with snapshot-based classification
• Evaluation: AUC, precision, recall, F1, cross-validation, DeLong's test
• Ablation Studies: Contribution of each pipeline component
• Backbone Comparisons: DenseNet, ViT-B/16, Swin-T, etc.

Interpretation: The full pipeline achieved the highest performance across all evaluation metrics, validating the importance of each module. These findings suggest that structural modeling and comprehensive spatial coverage are critical for detecting subtle FBA-related changes.

All P values were calculated using Fisher’s exact test.

Interpretation: The full pipeline achieved the highest performance across all evaluation metrics, validating the importance of each module. These findings suggest that structural modeling and comprehensive spatial coverage are critical for detecting subtle FBA-related changes.

In summary: This study presents an end-to-end DL pipeline combining high-precision airway segmentation and CT-based classification to detect radiolucent FBA. The model generalizes well across multiple cohorts and offers clinically meaningful improvements over expert interpretation in precision, highlighting its potential as a second-reader and triage aid in complex diagnostic settings.

This project is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0).

• 🧠 Project page: https://github.com/ZheChen1999/FBA_DL
• 📄 Full manuscript (preprint link if available)
• 📈 Supplementary tables and figures available in docs/

For detailed replication instructions, data access requests, or collaboration inquiries, please contact the corresponding authors.