Anand Devaraj

MD MBBS FRCR

Royal Brompton & Harefield NHS Foundation Trust

Consultant Radiologist

NLST (Learn more about this study)

NLST-751

Jan 14, 2021

Evaluation of Interstitial Lung Abnormalities using a deep learning developed algorithm

BACKGROUND
Interstitial lung abnormalities (ILAs) are findings on thoracic CT scans that are identified in individuals not suspected of having interstitial lung disease at the time of CT acquisition. ILAs may represent interstitial lung disease, or alternatively clinically insignificant findings.
ILAs are particularly prevalent in lung cancer screening cohorts, with a reported prevalence of up to 17% (Hoyer et al. 2018). It is known that the extent of ILAs on screening CT may progress over time , and also that ILAs are associated with an increased risk of mortality. However, there is limited understanding at present regarding i) which participants with ILAs will demonstrate progression and ii) how ILA progression is best captured. With an increasing emphasis on lung cancer screening implementation worldwide, there is a need for research to evaluate these two questions to understand how best to manage individuals in whom ILAs are discovered.

PROPOSED METHOD
We have developed a deep learning algorithm (Systematic Objective Fibrotic Imaging analysis Algorithm, SOFIA) for providing diagnostic support and outcome prediction in patients with established fibrotic lung disease (Walsh et al. 2018). We propose to test the ability of the SOFIA algorithm to a) identify participants with ILAs at risk of progression and b) demonstrate progression of ILA extent on serial CT. We will also evaluate the differences in the SOFIA algorithm outputs (fibrosis likelihood score) in CTs of participants with ILAs versus those with known interstitial lung disease and normal scans.
We would like to request baseline and follow up CTs from 3 categories of participants: 1) participants showing abnormalities classified as “reticular/reticulonodular opacities, honeycombing, fibrosis, scar” according to the “spiral CT abnormalities” data dictionary, 2) participants with a disease history diagnosis of “fibrosis of the lung” or “asbestosis” according to the “participant” data dictionary, and 3) 3000 participants without category 1 or 2 label as above.
Baseline CTs of participants from categories 1 and 2 will be reviewed visually by thoracic radiologists to confirm the presence of findings consistent with interstitial lung abnormalities and those without ILA will be excluded.
Thoracic radiologists will review scans where the SOFIA algorithm has identified ILAs from category 3 (normal scans) to confirm the presence or absence of ILAs. If ILAs are confirmed visually in this group, they will be included in the ILA cohort.
All CTs will ILA will be reviewed by thoracic radiologists to determine presence or absence of progression visually over time.
The SOFIA algorithm analysis will also be run on all CTs with ILA at baseline and follow up. We will validate the SOFIA scores of progression against visual scores. We will look for baseline and automated baseline CT characteristics that predict progression.

REFERENCES
Hoyer, N., W. Wille, L. H. Thomsen, T. et al. 2018. 'Interstitial lung abnormalities are associated with increased mortality in smokers', Respir Med, 136: 77-82.

Walsh, S. L. F., L. Calandriello, M. Silva, and N. Sverzellati. 2018. 'Deep learning for classifying fibrotic lung disease on high-resolution computed tomography: a case-cohort study', Lancet Respir Med, 6: 837-45.

To evaluate the performance of a deep learning algorithm prototype for characterization of interstitial lung abnormality (ILA) in participants of the NLST to include i) prediction of ILA progression, ii) automated demonstration of ILA progression.

To evaluate the differences in the deep learning algorithm outputs in CTs of participants with ILAs versus those with known interstitial lung disease and normal scans.