Deep learning in the diagnosis of idiopathic pulmonary fibrosis

<bold>Introduction:</bold> The diagnosis of idiopathic pulmonary fibrosis (IPF) mostly based on subjective data and requires multidisciplinary approach. Delay in diagnosing may negatively affect the prognosis. Deep learning algorithms to identify the UIP pattern in pulmonary fibrosis will contribute to early diagnosis and improve prognosis. <bold>Aims and objectives:</bold> We aim to standardize diagnosis of the UIP pattern in pulmonary fibrosis, to minimize diagnostic differences between clinicians, to increase the accuracy of diagnosis, to ensure early diagnosis. <bold>Methods:</bold> Between 2010 and 2023, 117 patients who were followed up in our outpatient clinic were included. Demographic data, functional evaluations and high-resolution thorax computed tomography (HRCT) scans were recorded. A database of 239 anonymised HRCT data consisting of a total of 73,522 images was created. Patients were evaluated by two different experienced thoracic radiologists and a multidisciplinary ILD council. HRCT images were categorized into 3 groups: typical UIP, probable UIP and alternative diagnosis. A deep learning algorithm was developed with CT images to classify the groups. <bold>Results:</bold> The algorithm achieved 99% and85% success rate for the training and the validation set. The model correctly predicted 4473 out of 5976 alternative diagnosis images (74.8%), 2002 out of 2064 probable UIP images (96.9%) and finally 5089 out of 6005 typical UIP images (84.7%) with 82% accuracy. The algorithm's precision for typical UIP, probable UIP and alternative diagnosis was 99%, 48% and 94%, and the recall was 85%, 97% and 75%, respectively. <bold>Conclusions:</bold> Deep learning algorithms may contribute the early diagnosis of IPF.