Artificial Intelligence Outperforms Radiologists in Classifying Fibrotic Lung Diseases

A deep learning algorithm more accurately diagnosed fibrotic lung diseases than a team of specialist thoracic radiologists in a case-cohort study published in The Lancet Respiratory Medicine. 

The algorithm and a team of 91 radiologists from multiple international thoracic imaging societies assessed a cohort of 150 high-resolution CT scans with fibrotic lung disease. The thoracic radiologists’ median accuracy was 70.7%, and the algorithm’s was 73.3%, outperforming two-thirds of the radiologists. The median interobserver agreement between the algorithm and the radiologist’s majority opinion (KW = 0.69) also outperformed that of 62% of the thoracic radiologists (KW = 0.67). 

#ERSCongress “#ILD: From bench to bedside”: Deep learning for classifying fibrotic lung disease on HRCT https://t.co/I497RJwPVO pic.twitter.com/lJnrb0VcoO

— JuanC Quijano-Campos💙 (@JCQuijanoCampos) September 16, 2018

The results of the study were also presented at the European Respiratory Society (ERS) meeting. 

“Just before someone asks, it’s not an attempt to replace radiologists,” Dr. Simon L.F. Walsh, one of the study’s authors, said during the presentation, according to MedPageToday. “That’s always the first question.” 

https://twitter.com/gburgeILDNN/status/1041235762933624837

Rather, the algorithm could serve as a diagnostic tool. Walsh also noted that idiopathic pulmonary fibrosis (IPF) patients may be limited if they live somewhere without top radiologists available. 

Researchers also tested the algorithm in differentiating between usual interstitial pneumonia and non-usual interstitial pneumonia and found its performance equal to that of the team of radiologists. Median interobserver agreement between the algorithm and radiologists was better than between the radiologists (KW = 0.64 vs. KW = 0.56). 

Podcast: Simon Walsh @slfwalsh @KingsCollegeLon discusses the potential for machine learning-assisted diagnosis of fibrotic #lung disease https://t.co/iRSRd4yTJg

Read the study presented at #ERSCongress: https://t.co/XGx9CZQ3Lf pic.twitter.com/T5xsX1GUY3

— The Lancet Respiratory Medicine (@LancetRespirMed) September 19, 2018

The study authors concluded, “High-resolution CT evaluation by a deep learning algorithm might provide low-cost, reproducible, near-instantaneous classification of fibrotic lung disease with human-level accuracy. These methods could be of benefit to centres at which thoracic imaging expertise is scarce, as well as for stratification of patients in clinical trials.” 

Read more about IPF diagnosis guidelines. 

Sources: The Lancet Respiratory Journal, MedPageToday