PO:32:187 | AI-assisted reading outperforms visual scoring of pulmonary CT scans to predict interstitial lung disease progression in systemic sclerosis

Francesca Motta¹, Antonio Tonutti¹, Federica Catapano², Gianluca Sellaro², Francesco Amati³, Anna Stainer³, Calogero Messana², Stefano Erba¹, Stefano Aliberti³, Marco Francone², Carlo Selmi¹, Maria De Santis¹ | ¹Humanitas University and IRCCS Humanitas Research Hospital, Department of Rheumatology and Clinical Immunology Rozzano, Milan, Italy; ²IRCCS Humanitas Research Hospital, Department of Radiology Rozzano, Milan, Italy; ³IRCCS Humanitas Research Hospital, Respiratory Unit Rozzano, Milan, Italy

Aim. Interstitial lung disease (ILD) is a major cause of morbidity and mortality in patients with systemic sclerosis (SSc) and the early detection of ILD progression is crucial for timely treatment to prevent poor outcome. However, reliable early markers of progression are currently lacking and treatment often starts when ILD has already progressed. We aimed to assess whether artificial intelligence (AI)-assisted analysis of chest high-resolution computed tomography (HRCT) is more sensitive than traditional visual scoring (1) in detecting early signs of ILD progression in patients with SSc.

Methods. In this retrospective, longitudinal study, 33 patients with SSc-ILD (79% female, 76% anti-Scl70+, 60% on treatment, of which 75% with mycophenolate at an average dose 2.075 g/day) underwent pulmonary function tests (PFT) and HRCT scans at two timepoints 12 months apart (T1, T2). A third timepoint (T3) after an additional 12 months was available in a subset of patients (n=20). HRCT scans were assessed by two chest radiologists for semi-quantitative evaluation of fibrosis and ground-glass opacities (GGO), and by AI-assisted software (Thoracic VCAR, GE Healthcare) to evaluate and quantify volume and percentage of fibrosis, GGO, emphysema, and consolidations. ILD progression was defined using Erice functional criteria (2) and the minimal clinically important worsening (MCIW) of FVC (decline of 3% or more over 12 months).

Results. Patient showed a significant reduction of forced vital capacity (FVC)% from T1 to T2 (85 –SD 21- to 80 – SD 17- %, p=0.04), but no difference in other PFT parameters. Both visual and AI-assisted HRCT analyses showed significant increase in fibrosis scores over time (visual: 1.5 to 1.6, p=0.0005; AI: 2.2 to 2.4%, p=0.0007). AI also detected a significant increase in GGO% (15.5 to 15.8%, p = 0.04) and reduction in normal lung (75.6 to 73%, p=0.009). AI-derived HRCT parameters strongly correlated with PFT values, particularly FVC and diffusing capacity for carbon monoxide (DLCO), at both T1 and T2. In contrast, visual scores showed weaker correlations (Tables 1 and 2). Concordance between visual and AI assessments of GGO and fibrosis was low or absent at both timepoints. Progressive ILD at T3 was predicted by the increase of AI-determined volume of GGO from T1 to T2, even after adjustment for age, gender, and disease duration (P=0.04, OR 9, 95%CI 1.37-15). Visual scoring did not predict progression.

Conclusions. The AI-assisted quantitative HRCT analysis detects early SSc-ILD progression and may identify patients at risk of functional decline better and earlier than PFT.

References
1. Kazerooni EA, et al. AJR Am J Roentgenol. 1997 Oct;169(4):977-83. 2. George PM, et al. Lancet Respir Med. 2020 Sep;8(9):925-934.