The following is a summary of “Development and validation of a CT-based deep learning algorithm to augment non-invasive diagnosis of idiopathic pulmonary fibrosis,” published in the NOVEMBER 2023 issue of Pulmonology by Maddali, et al.
For a study, researchers sought to enhance the non-invasive diagnosis of idiopathic pulmonary fibrosis (IPF) by developing and validating a machine learning algorithm utilizing computed tomography (CT) scans exclusively. The objective was to address the uncertainty in identifying the radiologic usual interstitial pneumonia (UIP) pattern, which often necessitates invasive surgical biopsy.
The primary methodology involved the development of a deep learning convolutional neural network (CNN) designed exclusively for processing CT images. The algorithm aimed to predict the presence of idiopathic pulmonary fibrosis (IPF) within the spectrum of interstitial lung diseases (ILDs), utilizing multidisciplinary discussion (MDD) consensus diagnosis as the reference standard. The training process utilized a comprehensive multi-center dataset comprising over 2000 ILD cases. To refine and optimize the algorithm, a US-based multi-site cohort consisting of 295 cases was employed for tuning. External validation was conducted using an independent dataset sourced from European and South American origins, encompassing 295 cases.
In the tuning set, the model exhibited robust discriminative ability, with an area under the receiver operating characteristic curve (AUC) of 0.87 (95% CI: 0.83–0.92), effectively distinguishing idiopathic pulmonary fibrosis (IPF) from other interstitial lung diseases (ILDs). The sensitivity and specificity were calculated at 0.67 (95% CI: 0.57–0.76) and 0.90 (95% CI: 0.83–0.95), respectively. In stark contrast, assessments conducted before multidisciplinary discussion (MDD) diagnosis displayed lower sensitivity at 0.31 (95% CI: 0.23–0.42) and comparable specificity at 0.92 (95% CI: 0.87–0.95). External validation in a separate test set also demonstrated a c-statistic of 0.87 (95% CI: 0.83–0.91). The model’s consistent performance extended across diverse CT scanner manufacturers and various slice thicknesses.
The deep learning algorithm, relying solely on CT images, consistently excelled in identifying IPF among ILD cases. Its robust performance, with a focus on the radiologic UIP pattern, suggested promising generalizability across different CT equipment, emphasizing its potential as a valuable diagnostic tool in IPF assessment.
Source: resmedjournal.com/article/S0954-6111(23)00316-5/fulltext