# Automated scoring of airway abnormalities and mucus plugging in chest magnetic resonance imaging of cystic fibrosis using artificial intelligence

## TL;DR

This paper introduces an AI method to automatically detect lung abnormalities in cystic fibrosis using MRI, reducing the need for manual analysis and radiation.

## Contribution

A deep learning approach for automated classification of bronchiectasis and mucus plugging in chest MRI for cystic fibrosis.

## Key findings

- The model achieved substantial agreement for mucus plugging detection (κ = 0.68) and strong discrimination (AUROC = 0.90).
- The method showed moderate agreement for bronchiectasis/wall thickening (κ = 0.53) with strong discrimination (AUROC = 0.87).
- Grad-CAM analyses confirmed the model's attention aligned with relevant pathologies, and it generalized well to an external dataset.

## Abstract

Cystic fibrosis is characterized by progressive lung damage, requiring life-long medical treatment and monitoring. This emphasizes the need for reliable, radiation-free imaging and automated analysis of lung disease activity. We present a deep learning-based approach for classifying two key pathologies, bronchiectasis/wall thickening and mucus plugging, on T2-weighted chest MRI. Retrospectively, 627 MRI scans from 164 patients (mean age 7.0 ± 6.2 years; range 0.1–53.0 years) were collected. Chest MRI were preprocessed with an nnU-Net to segment lung halves, followed by an atlas-based lung lobe approximation. Leveraging a dual-view architecture processing coronal and axial slices, our approach addresses limitations inherent in manual scoring, such as reader variability and substantial labor requirements. We evaluated a single model trained on all lobes and models specialized for each lobe. Cross-validation revealed substantial agreement for mucus plugging (κ = 0.68) with strong discrimination (macro AUROC = 0.90) and excellent reliability (Pearson’s r = 0.84). For bronchiectasis/wall thickening, agreement was moderate (κ = 0.53) but discrimination remained strong (macro AUROC = 0.87), with Pearson’s r = 0.74. The mean differences and 95 % limits of agreement for both pathologies aligned closely with the reader variability previously reported. Grad-CAM analyses demonstrated spatial alignment of model attention with relevant pathologies, and external testing on ten patients from an independent centre confirmed promising generalization. These findings represent a significant step toward automated MRI-based assessment for CF-related lung changes. Extending the approach to additional MRI scoring items may also improve granularity and clinical applicability, ultimately aiding in more personalized CF management.