Deep Learning Quantitative Computed Tomography Analysis in the Australian Idiopathic Pulmonary Fibrosis Registry

Abstract Rationale: Idiopathic pulmonary fibrosis (IPF) is a universally progressive lung disease, however individualised prognostication remains challenging. A number of increasingly sophisticated automated algorithms have been applied to computed tomography in people with IPF, demonstrating prognostic value. Methods: Previously validated deep learning-based algorithms were applied to baseline CT scans from participants of the Australian IPF Registry (AIPFR) to quantify imaging biomarkers including lung volume (Lung8), and percentages of lung volume containing airways (Air8), vasculature (Vascul8), and fibrosis (Fibr8). Associations of imaging biomarkers with baseline lung function, disease progression at 12 months, and transplant-free survival were assessed. Further analysis was performed stratified by background anti-fibrotic therapy and disease severity. Results: All four algorithms were successfully applied to CT scans from 282/515 participants of the AIPFR (mean age 70.7 years, 66.7% male). Mean FVC% was 80% and mean DLCO% was 49%. Deep-learning analysis revealed a mean lung volume of 3.8±1.0L, airway volume of 6.5±2.8%, vascular volume of 2.8±5.7%, and fibrosis volume of 17.3±11.7%. Volumes of fibrosis, airways and vasculature were negatively correlated with FVC% (r=-0.50, p<0.001, r=-0.20, p<0.001, r=-0.45, p<0.001 respectively) and DLCO% (r=-0.43, p<0.001, r=-0.11, p=0.09, r=-0.26, p<0.001 respectively). In univariable analysis, all imaging biomarkers were significantly associated with reduced transplant-free survival (Lung volume: HR 1.00, 95%CI 1.00-1.00, p=0.01; Airway volume: HR 3.75, 95%CI 2.20-6.40, p<0.001; Vascular volume: HR 2.28, 95%CI 1.82-2.84, p<0.001; Fibrosis volume: HR 1.04, 95%CI 1.03-1.05, p<0.001). In multivariable analysis of all imaging biomarkers, fibrosis volume (HR 1.03, 95%CI 1.02-1.05, p<0.001) and airway volume (HR 3.16, 95%CI 1.86-5.37, p<0.001) remained independently associated with transplant-free survival. These findings remained true for progression at 12 months and for anti-fibrotic treated patients (n=94). In patients with an FVC <80% larger airway volume was associated with worse transplant-free survival, whereas in those with an FVC >=80% increased fibrosis volume predicted poorer transplant-free survival. Conclusions: Automated deep learning-based CT analysis provides robust, quantitative imaging biomarkers that correlate with lung function and independently predict outcomes in IPF. Prognostic significance of imaging biomarkers varies with disease severity, indicating that extent of traction bronchiectasis may be of more relevance in advanced disease whilst extent of fibrosis may be more relevant in early disease.