Development and validation of a deep learning-based algorithm for quantifying bronchiolitis obliterans in pediatric computed tomography.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Seoul National University Hospital, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Department of Radiology, Pusan National University Yangsan Hospital, Pusan National University College of Medicine, Busan, Republic of Korea.
- Institute of Radiation Medicine, Seoul National University Medical Research Center, Seoul, Republic of Korea.
- Department of Pediatrics, Seoul National University Hospital, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Innovative Medical Technology Research Institute, Seoul National University Hospital, Seoul, Republic of Korea.
Abstract
To develop and validate a deep learning-based algorithm for quantifying bronchiolitis obliterans (BO) on pediatric chest CT. This retrospective study included 86 children (39 males; median age, 10 years) diagnosed with BO who underwent both inspiratory and expiratory CT between January 2018 and November 2021. The deep learning-based BO quantification model was trained on 26 CT scans using a 3D nnU-Net, with radiologist-segmented low attenuation regions (LARs) serving as ground truth. Model performance was evaluated through internal test with 4 CT scans and external test with 6 CT scans. Intra-vendor robustness was assessed using 22 CT scans with varying reconstruction methods, kernel types, and slice thicknesses. Comparison with semi-quantitative radiologist grading was performed using 28 CT scans. Dice similarity coefficient (DSC), sensitivity, and precision were used to evaluate model performance. The model achieved a DSC of 85.41 ± 3.28%, sensitivity of 85.14 ± 7.66%, and precision of 86.21 ± 3.92% in the internal test, and 82.53 ± 4.34%, 82.17 ± 6.15%, and 84.15 ± 3.16% in the external test, respectively. For intra-vendor robustness, no significant differences in BO quantification were observed across different reconstruction methods, kernel types, and slice thicknesses (all p > 0.05). Compared to radiologists' grading, the model demonstrated strong to very strong correlations across all lung lobes (all p < 0.001). The model demonstrated accurate quantification of BO on pediatric CT, with good agreement with the radiologist-segmented ground truth. This study presents a 3D nnU-Net-based deep learning algorithm for robust quantification of BO on pediatric CT, providing reproducible measurements of LARs.