Op0167 Utility of Novel Computer Based Deep-Learning Analysis for Patients With Connective Tissue Disease Related Interstitial Lung Disease Compared to Conventional Visual Scoring Method

Background: Quantifying the extent of interstitial lung disease (ILD) on high-resolution computed tomography (HRCT) by conventional visual scoring method (CVSM) is limited by objectivity and reproducibility. Objectives: To investigate the utility of novel computer based deep-learning analysis (CBDLA, QZIP-ILD, Ziosoft, Tokyo) compared to CVSM for evaluating connective tissue disease related ILD (CTD-ILD). Methods: Patients with CTD-ILD visiting our hospital between 2020 to 2023 were consecutively enrolled. The lesion ratios (i.e., ratio to the total lung field) for ground glass opacity (GGO), consolidation and total fibrotic lesion (TFL) on CBDLA were compared to CVSM using Spearman's correlation and Bland-Altman analysis. TFL was defined by the sum of consolidation with fibrosis, honeycombing, reticulation and tractionbronchiectasis on CBDLA, whereas that of honeycombing and reticulation on CVSM. The relationship between forced vital capacity (FVC)% predicted and TFL ratio on CBDLA or CVSM were also investigated. Results: Seventy-three patients (median age, 66 years, 13 men) were included. Disease subtype was as follows: 21 (28.7%) patients with rheumatoid arthritis (RA), 27 (36.9%) with systemic sclerosis (SSc), 22 (30.1%) with myositis, 10 (13.6%) with antineutrophil cytoplasmic antibody associated vasculitis (AAV). The median lesion ratios for GGO, consolidation and TFL were 3.63%, 0.69% and 2.38% on CBDLA whereas 3.21%, 0.00% and 4.16% on CVSM (Table 1). The lesion ratios for GGO, consolidation and TFL on CBDLA was significantly correlated with that on CVSM (r = 0.671, 0.456 and 0.788, P <0.05). Significant high correlations for GGO ratios and TFL ratios between CBDLA and CVSM were constantly observed irrespective of disease subtype (r = 0.496 in RA to 0.847 in SSc for GGO, r = 0.663 in SSc to 0.903 in AAV for TFL). The mean differences of GGO, consolidation and TFL was -0.31, 0.76 and -4.01%, which were within 5%. The correlation between the value of FVC% predicted and TFL ratio was moderate in CBDLA (r = -0.529, P <0.01), whereas fair in CVSM (r = -0.234, P = 0.045). Performance of TFL ratio for FVC% predicted <70% using CBDLA was significantly higher than CVSM (area under the receiver operating characteristic curve 0.779, 0.583, DeLong's test, P <0.05). Conclusion: In conclusion, the TFL ratios on CBDLA showed strong correlation with CVSM and better performance for predicting physiologically severe CTD-ILD. This new CBDLA not only functions as an alternative to CVSM but is also expected to be used in clinical practice as a means to more accurately evaluate CTD-ILD. REFERENCES: NIL. Acknowledgements: NIL. Disclosure of Interests: None declared.