Scientific Session 24 — SS24: Chest Imaging - Pulmonary Embolism, Interstitial Lung Diseases and Applications of Deep LearningThursday, May 9, 2019
2531. Using Deep Learning to Predict Severity of Restrictive Pulmonary Function From Chest Radiographs of Patients With Interstitial Lung Disease
Chan J*, Lanfredi R, Tasdizen T, Srikumar V, Schroeder J. University of Utah, Salt Lake City, UT
Address correspondence to J. Chan (firstname.lastname@example.org)
Objective: We aim to predict the severity of restrictive pulmonary function in patients with interstitial lung disease (ILD) from chest radiographs by using deep convolutional neural networks. Restricted pulmonary function is directly attributable to changes in the lung structure, such as alveoli wall thickening, and many studies have found an association between morphologic features of pulmonary fibrosis on chest CT with pulmonary function. Although these morphologic features are readily visible on chest CT, they are less conspicuous and difficult to quantify on chest radiography for a radiologist; however, multiple recent studies have demonstrated the success of deep-learning algorithms, a type of machine learning, to detect subtle abnormalities on chest radiography. We hypothesize that a deep-learning convolutional neural network (CNN) can predict pulmonary function from chest radiographs of patients’ with interstitial lung disease, a topic not previously studied in the literature.
Materials and Methods: This study is IRB approved. We have a dataset of approximately 2000 frontal chest radiographs of adult patients with interstitial lung disease, acquired and interpreted by the Department of Radiology within our institution, with associated clinical parameters, deidentified radiology reports, and pulmonary function test (PFT) data. Cases will be chosen at random to populate the training, validation, and testing groups, and parameters of the convolutional neural network including number of layers, filter size, and number of neurons per layer will be optimized on the validation set in collaboration with our colleagues at the University of Utah Scientific Computing and Imaging Institute. Deep-learning algorithms will be trained, validated, and tested for each PFT component of interest. We will use receiver operator curves and area under the curve to evaluate model performance for each PFT component.
Results: We have successfully transferred the chest radiographs, clinical parameters, PFT data, and deidentified reports into our HIPPA-compliant center for high-performance computing (CHPC). In collaboration with our computer science colleagues, we have begun training a CNN on a test patient population.
Conclusion: The added pulmonary function information on baseline and subsequent chest radiographs from our CNN will aid the radiologist and pulmonologist in tracking disease progression and providing prognostic information. The quantification of pulmonary function from chest radiography, a topic not previously discussed in the literature, will also dramatically increase the specificity and timeliness of care of patients with lung disease, not limited to pulmonary fibrosis.