Deep-learning CT imaging algorithm to detect usual interstitial pneumonia pattern in patients with systemic sclerosis-associated interstitial lung disease: association with disease progression and survival

Objectives: Interstitial lung disease (ILD) is the most common cause of death in patients with systemic sclerosis (SSc), although disease behavior is highly heterogeneous. While a usual interstitial pneumonia (UIP) pattern is associated with worse survival in other ILDs, its significance in SSc-ILD is unclear. We sought to assess the prognostic utility of a deep-learning HRCT algorithm of UIP probability in SSc-ILD.

Methods: Patients with SSc-ILD were included if HRCT images, concomitant lung function tests, and follow-up data were available. We used the Systematic Objective Fibrotic Imaging analysis Algorithm (SOFIA), a convolution neural network algorithm which provides probabilities of a UIP pattern on HRCT images. These were converted into the Prospective Investigation of Pulmonary Embolism Diagnosis (PIOPED)-based UIP probability categories. Decline in lung function was assessed by mixed-effect model analysis and relationship with survival by Cox proportional hazards analysis.

Results: 522 patients were included in the study. 19.5% were classified as UIP not in the differential, 53.5% as low probability of UIP, 25.7% as intermediate probability of UIP, and 1.3% as high probability of UIP. A higher likelihood of UIP probability expressed as PIOPED categories was associated with worse baseline FVC, as well as with decline in FVC (p= 0.008), and worse 15-year survival (p= 0.001), both independently of age, gender, ethnicity, smoking history, and baseline FVC or Goh et al. staging system.

Conclusion: A higher probability of a SOFIA-determined UIP pattern is associated with more advanced ILD, disease progression, and worse survival, suggesting that it may be a useful prognostic marker in SSc-ILD.