New research suggests that a deep learning classification tool may significantly enhance the diagnosis and classification of usual interstitial pneumonia (UIP) on chest computed tomography (CT) scans.
For the retrospective study, recently published in Chest, researchers assessed the use of a deep learning classification modality, trained on 1,934 chest CTs, for predicting UIP in an independent performance cohort of 565 patients and a multicenter cohort of 265 patients with interstitial lung disease (ILD).
Noting that the radiologist assessed visual UIP prevalence rate was 12.4 percent in the performance cohort, the researchers found that the deep learning classification tool had a 93 percent sensitivity rate and an 86 percent specificity rate in this cohort for discriminating the radiologist-determined ground truth UIP. For the ILD cohort, radiologists noted a 37.1 percent prevalence of visual UIP, and the deep learning classification tool demonstrated an 81 percent sensitivity and a 77 percent specificity rate for UIP.
The study authors also noted a 66 percent positive predictive value (PPV) and a negative predictive value (NPV) of 95 percent for the deep learning classification tool in the performance cohort. However, in the ILD cohort, the deep learning tool had an 85 percent PPV and a 70 percent NPV, according to the study.
“These findings suggest this (deep learning) UIP classifier predicts radiologist determined ground truth UIP with good test performance across a wide range of UIP prevalence,” wrote lead study author Jonathan H. Chung, M.D., chief of the Section of Cardiopulmonary Imaging and vice-chief of quality in the Department of Radiology at the University of Chicago Medicine, and colleagues.
Three Key Takeaways
- Enhanced diagnostic accuracy. The deep learning classification tool showed promising results in enhancing the diagnosis and classification of UIP. It exhibited a high sensitivity and specificity for identifying UIP in chest CT scans.
- Potential clinical application. The tool may be valuable for healthcare facilities that lack specialized thoracic radiologists, as it can provide a reliable and efficient method for identifying UIP, which is often associated with idiopathic pulmonary fibrosis (IPF) and other interstitial lung diseases (ILDs).
- Study limitations. The study authors acknowledge limitations, including the low positive predictive value due to the rare nature of ILD, and concede that the tool's performance with less typical UIP findings on CT remains unclear.
While noting that UIP suggests the development of idiopathic pulmonary fibrosis (IPF) without a clear etiology for ILD, the study authors pointed out that UIP has also been linked to fibrotic hypersensitivity pneumonitis, patients with asbestos and ILD caused by autoimmune conditions such as rheumatoid arthritis. Accordingly, the researchers said the deep learning classification tool for UIP could be beneficial in this patient population.
“An automated, CT-based tool that quickly and reliably identifies UIP would be of high value to the field as this would allow for more confident application of ILD diagnostic criteria, especially for centers without a thoracic radiologist,” added Chung and colleagues. “Such a modality could also allow for systematic screening for IPF and other forms of fibrosing ILDs with this high-risk morphology.”
(Editor’s note: For related content, see “COVID-19 and Cancer: What a New Chest CT Study Reveals,” “Study Finds Over Half of Patients with COVID-19 Pneumonia Have Pulmonary Abnormalities One Year Later” and “Chest CT Study Shows Benefits of COVID-19 Vaccines in Reducing Incidence and Severity of Related Pneumonia.”)
In regard to study limitations, the authors point out that any tool geared toward the detection and classification of ILD will have a low positive predictive value due to the rare nature of the disease. The researchers acknowledged that a significant portion of the study cohort was comprised of participants from clinical trials for IPF who would have had a higher likelihood of classic UIP presentation on CT. Accordingly, they noted a lack of clarity as to the performance of the deep learning tool for recognizing less typical findings of UIP on CT.
