Abstract
BACKGROUND: In patients with idiopathic pulmonary fibrosis (IPF), computed tomography (CT) quantification using artificial intelligence (AI) has been explored as a method to assess the therapeutic response to antifibrotic agents; however, studies evaluating long-term follow-up outcomes remain scarce. We investigated AI-driven quantitative analysis for long-term follow-up chest CT of IPF patients undergoing pirfenidone treatment. METHODS: Among the 2,223 patients diagnosed with interstitial lung disease by chest CT at Jeonbuk National University Hospital, 36 patients with a multidisciplinary diagnosis of IPF were included in the study after excluding those who had not undergone surgical lung biopsy or did not have available pulmonary function tests (PFTs). These 36 patients underwent high-resolution computed tomography (HRCT) along with concurrent PFTs over a 10-year period and were categorized into two groups: those treated with pirfenidone (n=17) and those not treated with pirfenidone (n=19). Quantitative texture analysis was performed using a deep convolutional neural network to calculate fibrotic scores, defined as the combined mean percentage of two fibrotic components-reticulation and honeycombing, with or without accompanying ground-glass opacities-across the entire lung. This analysis aimed to assess treatment response in IPF patients receiving pirfenidone. Repeated measures analysis of variance (ANOVA) was used to evaluate the correlation between changes in pulmonary function and fibrotic scores over time. RESULTS: The final study population comprised 36 patients, with a mean age of 67.1±7.7 years. Patients (DLCO: 58.0%±21.0%) who received pirfenidone (n=17) exhibited lower DLCO values at the final follow-up compared to the untreated group (n=19) (DLCO: 69.0%±21.7%), although the difference was not statistically significant (P=0.260). However, in the treated group (n=17), patients with progression despite pirfenidone treatment (n=6) (fibrotic score: 27.1%±12.1%) showed a markedly greater increase in mean AI fibrotic scores at the final follow-up compared to those with no or little change (n=11) (fibrotic score: 10.9%±8.7%), with the difference approaching statistical significance (P=0.076). There was a significant correlation between the decrease in DLCO values and the increase in AI fibrotic score in patients with pirfenidone on long-term follow-up (P<0.01). CONCLUSIONS: AI-driven quantitative analysis of HRCT images in patients with IPF enables objective monitoring of the effects of pirfenidone on the progression of pulmonary fibrosis on long-term follow-up.