Quantitative Chest CT in COPD: Can Deep Learning Enable the Transition?

See also article by Hasenstab et al in this issue.

Mannudeep K. Kalra, MD, is an attending thoracic radiologist at Massachusetts General Hospital, director of MGH Webster Center for Quality and Safety, and professor of radiology with the Harvard Medical School. His research interests include CT technology assessment, radiation dose optimization, and deep learning applications in cardiothoracic imaging.

Shadi Ebrahimian, MD, is a postdoctoral research fellow in the department of radiology at Massachusetts General Hospital. Her main research areas of interest are advanced CT reconstruction techniques and deep learning algorithms in thoracic imaging.

With a variable prevalence of 5%–20% among U.S. adults and a total annual cost estimate of $36 billion per year, the impact of chronic obstructive pulmonary disease (COPD) on health care is substantial in both the United States and worldwide, where it affects close to 200 million people (1). Spirometry-based pulmonary function tests are the mainstay for establishing the diagnosis, classifying disease severity, assessing treatment response, and predicting mortality and outcomes in patients with COPD (2). However, spirometry can miss active small airway diseases and has a poor correlation with clinical outcomes such as exacerbation frequency and treatment response (3). Although chest CT holds an important place in the diagnosis and management of COPD, the lack of identification of early local bronchial inflammation, subjective nature of clinical reporting, and associated radiation dose are significant limitations.

In the context of subjective reporting, Hasenstab and colleagues present compelling evidence of how deep learning (DL) can enhance the value of chest CT by enabling a generalizable, and more importantly, an easily executable way of obtaining quantitative CT information on low-attenuating areas and air trapping in patients with COPD (4). The authors developed and tested a DL algorithm to stage disease severity and predict progression and mortality from data sets derived from their institution, the National Lung Screening Trial, and the COPDGene study. The DL algorithm enabled automatic assessment of the percentage of emphysema (percentage of lungs ≤ −950 HU on inspiratory phase images) and total lung involvement (percentage of lungs with either emphysema or air trapping [identified as ≤ 100 HU difference between inspiration and expiration phase images in lung regions > −950 HU at inspiration]). On the basis of the percentage of emphysema and/or total lung involvement, the authors proposed five CT-based COPD stages, including normal, mild, moderate, severe, and very severe. The proposed CT-based COPD stages predicted spirometry-based Global Initiative for Chronic Obstructive Lung Disease (GOLD) criteria with high areas under the curve (0.86–0.96). Furthermore, the proposed CT-based COPD stages both with and without GOLD criteria could successfully predict disease progression (odds ratio, 1.50–2.67) and mortality (hazard ratio, 2.23) (P < .001 for both outcome measures).

The strengths of their work lie in DL-based automation of quantitative emphysema and air trapping evaluation at chest CT, the robust training and test data sets from different time points, health care sites, and scanner types, and robust clinical and laboratory biomarkers as well as outcome predictors (disease severity and mortality). Generalizability and reproducibility are essential to successfully deploy DL algorithms, particularly in modifying treatment options based on disease severity, progression, and mortality. The inclusion of CT data from at least three sources in the study (4) helps establish the assessed DL algorithm’s generalizability across multiple scanners and different patient types. The study (4) is, however, not without some limitations. The authors (4) acknowledge that their algorithm did not evaluate bronchial inflammation or thickening (inferred from bronchial luminal dimensions, wall thickness, and their ratios), which could confound or enhance disease quantification and prognostication. A sizeable group of patients with obstructive airway diseases (with decreased percentage predictive forced expiratory volume in 1 second [FEV₁pp] < 80%; reduced forced vital capacity [FVC]) and preserved FEV₁/FVC ratio (> 0.7), classified as preserved ratio impaired spirometry, were also excluded from the study (4). Another limitation of their work pertains to lack of verification of their algorithm under suboptimal conditions related to the scanner (such as those related to artifacts or lower radiation dose < 50 mAs) and/or patient (inadequate full inspiration or expiration, very large patient body habitus, and other coexisting diseases such as asthma, bronchiectasis, pulmonary fibrosis). Suboptimal expiration and respiratory motion artifacts are not infrequent in patients with severe or very severe diseases or those experiencing acute exacerbations or complications. There is a need to identify and exclude or normalize CT data extracted under such circumstances.

Despite the stated limitations, results obtained by Hasenstab et al (4) are well-supported by a sufficient body of evidence and publications on quantitative CT in patients with COPD that document the ability of DL-based quantitative algorithms to detect and quantify emphysema and air trapping at chest CT and to predict outcomes such as disease symptoms, progression, and mortality (5–8). According to the American College of Radiology Data Science Institute, the U.S. Food and Drug Administration (FDA) has cleared several DL-based algorithms for quantitative lung analysis in patients with COPD (9). As reported in Hasenstab study (4), some algorithms provide the whole lung, per lung, and lung lobe level distribution of emphysema, air trapping, and bronchial dimensions.

Technological advances such as improved scanner efficiency and better iterative and DL image reconstruction techniques can now deliver images with low noise and artifact content at submillisievert doses. Wide-detector array CT scanners and those with high pitch and fast rotation time can scan the entire chest in as little as one-half second and help reduce the frequency and severity of motion artifacts. Initial patient studies with ultra-high-resolution images obtained from photon-counting detector CT suggest gains in spatial resolution (the prospect of visualizing and quantifying small airways < 2 mm), improved stability of CT attenuation scale (more stable CT attenuation threshold), and perhaps hope for targeted radiotracer markers of inflammation (for patients with airway inflammation) (10). These developments may improve the automatic quantification and prognostication of disease burden in patients with COPD using DL-enabled quantitative chest CT.

Availability of FDA-cleared DL algorithms (9) and publications such as those from Hasenstab et al (4) and others (5–8) can help transition qualitative semantic reports of chest CT in patients with COPD to a hybrid format where DL adds quantitative information on the extent of emphysema, air trapping, bronchial wall thickening, and a CT-based COPD staging. Such transition can hopefully spur or follow further prospective studies in patients with COPD to address the limitations of current publications (4) and include a spectrum of confounding or compounding pulmonary diseases (such as bronchial asthma and bronchiectasis). We hope that DL algorithms can reduce or avoid the burden of manual image processing and incorporation of quantitative information into radiology report text. The promising array of DL algorithms can help optimize treatment strategies with a better classification of imaging and spirometry-based disease staging, severity, and prognostication.