Deep learning-based quantification of traction bronchiectasis severity for predicting outcome in idiopathic pulmonary fibrosis

Abstract

Aim: We investigated the prognostic utility a novel deep learning algorithm for quantifying severity of traction bronchiectasis in patients with idiopathic pulmonary fibrosis (IPF) enrolled in the Australian IPF Registry (AIPFR). Methods: Visual evaluation of HRCTs from the AIPFR was performed by 2 expert thoracic radiologists evaluated. Total airway volume (TAV) was quantified using a novel 3D U-Net-based deep learning algorithm. SOFIA UIP probability scores were obtained using a previously reported deep learning algorithm, trained in the identification of UIP features. Results: Total airway volume was an independent predictor of mortality when controlling for visual-based evaluation of total fibrosis extent (HR 1.96, p<0.0001), %Predicted FVC (HR 2.15, p<0.0001) or the CPI (n=217, HR 1.52, p=0.02. On bivariable analysis both TAV (HR 2.13, p<0.0001) and SOFIA-UIP probability (HR 1.30, p<0.0001) independently predicted mortality. On bivariable analysis with total fibrosis extent, TAV independently predicted mortality in UIP-like disease (HR 1.50, p=0.03) and was the only predictor of mortality (HR 5.33, p<0.0001) in those meeting indeterminate/alternative diagnosis criteria. An increase in TAV of 1% of total lung volume was associated with a 3-fold increased likelihood of developing progressive disease (OR 3.04 p=0.009) when controlling for total fibrosis extent. Conclusion: In IPF, automated quantification of TAV predicts mortality independently of total fibrosis extent on HRCT and can be used to identify patients at risk of progression at 12 months. In collaboration with the AIPFR and The Open Source Imaging Consortium