Harnessing deep learning to detect bronchiolitis obliterans syndrome from chest CT.

backgroundBronchiolitis Obliterans Syndrome (BOS), a fibrotic airway disease that may develop after lung transplantation, conventionally relies on pulmonary function tests (PFTs) for diagnosis due to limitations of CT imaging. Deep neural networks (DNNs) have not previously been used for BOS detection. This study aims to train a DNN to detect BOS in CT scans using an approach tailored for low-data scenarios.

methodsWe trained a DNN to detect BOS in CT scans using a co-training method designed to enhance performance in low-data environments. Our method employs an auxiliary task that makes the DNN more sensitive to disease manifestations and less sensitive to the patient's anatomical features. The DNN was tasked with predicting the sequence of two CT scans taken from the same BOS patient at least six months apart. We evaluated this approach on CT scans from 75 post-transplant patients, including 26 with BOS, and used a ROC-AUC metric to assess performance.

resultsWe show that our DNN method achieves a ROC-AUC of 0.90 (95% CI: 0.840-0.953) in distinguishing BOS from non-BOS in CT scans. Performance correlates with BOS progression, with ROC-AUC values of 0.88 for stage I, 0.91 for stage II, and 0.94 for stage III BOS. Notably, the DNN shows comparable performance on standard- and high-resolution CT scans. It also demonstrates the ability to predict BOS in at-risk patients (FEV1 between 80% and 90% of best FEV1) with a ROC-AUC of 0.87 (95% CI: 0.735-0.974). Using visual interpretation techniques for DNNs, we reveal sensitivity to hyperlucent/hypoattenuated areas indicative of air-trapping or bronchiectasis.

conclusionsOur approach shows potential for improving BOS diagnosis by enabling early detection and management. The ability to detect BOS from standard-resolution scans at any stage of respiration makes this method more accessible than previous approaches. Additionally, our findings highlight that techniques to limit overfitting are crucial for unlocking the potential of DNNs in low-data settings, which could assist clinicians in BOS studies with limited patient data.