Development and validation of a deep learning-based algorithm for quantifying bronchiolitis obliterans in paediatric...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Pediatric Radiology
• Pulmonary Disease Quantification

Background:

• Bronchiolitis obliterans (BO) is a serious complication following pediatric lung transplantation and other conditions.
• Accurate quantification of BO on CT scans is crucial for patient management and treatment assessment.
• Current quantification methods can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a deep learning (DL) algorithm for objective quantification of BO on pediatric chest CT.
• To assess the robustness and accuracy of the DL model against radiologist segmentations and grading.

Main Methods:

• A retrospective study utilized 86 pediatric chest CT scans from patients diagnosed with BO.
• A 3D nnU-Net deep learning model was trained on radiologist-segmented low attenuation regions (LARs) as ground truth.
• Model performance was evaluated using internal and external test sets, intra-vendor robustness testing, and comparison with semi-quantitative radiologist grading.

Main Results:

• The DL model achieved high performance metrics, including Dice Similarity Coefficients (DSC) of 85.41% (internal) and 82.53% (external).
• The algorithm demonstrated robustness across varying CT acquisition parameters (reconstruction methods, kernel types, slice thicknesses).
• Strong to very strong correlations were observed between the DL model's quantification and radiologist grading across all lung lobes (p < 0.001).

Conclusions:

• The developed 3D nnU-Net deep learning algorithm provides accurate and reproducible quantification of BO on pediatric CT.
• The model shows good agreement with radiologist-segmented ground truth, offering a reliable tool for clinical assessment.
• This AI-driven approach enhances the objective measurement of low attenuation regions (LARs) in pediatric BO.