Do You Believe in Magic? Computed Tomography, Artificial Intelligence, and Chronic Obstructive Pulmonary Disease

Released in 1965, the song “Do You Believe in Magic” was the first hit record of the folk-rock band the Lovin’ Spoonful. Although the song reflected the optimism of that distant time and a hope that music would magically lead people to a better world, the title seems strangely relevant to medicine today. The reason for this lies in an article published 3 years earlier by the science fiction author Arthur C. Clarke, best known today as the writer of 2001: A Space Odyssey. Clarke formulated three “laws” for the future, the best remembered being the third: Any sufficiently advanced technology is indistinguishable from magic (1). That view would certainly seem true for the imaging of emphysema.

In the mid-1960s, emphysema could only be conclusively diagnosed at autopsy, although plain chest X-rays could identify severe disease associated with hyperinflation, reduced pulmonary vascularity, or large bullae (2). This changed in the mid-1980s, when computed tomography (CT) scanning provided the first, initially rather grainy, high-resolution pictures of lung structure in obstructive lung disease (3). The subsequent decades have seen astonishing progress, beginning with the separation of centriacinar and panacinar disease in vivo and then the use of density masks to quantify the relative proportion of affected lung (4). More quantitative data followed, with a major role played by investigators in the COPDGene and SPIROMICS projects showing that CT-defined emphysema was related to the progression of chronic obstructive pulmonary disease (COPD) and the risk of exacerbation (5–7). The ability to manipulate the digitized image within a computer allowed parametric response mapping to identify areas of lung where microscopic, small airway damage and emphysema are present (8). As a result, it is clear that lung damage occurs early in the natural history of smoking-related COPD (9). However, to determine the presence of microscopic disease requires voxel-to-voxel matching of inspiratory and expiratory scans with an attendant increase in scanning time, protocol complexity, and radiation dose.

In this issue of the Journal, Chaudhary and colleagues (pp. 1185–1195) report a seemingly magical way of overcoming this problem (10). They applied their recently developed method of creating a virtual expiratory CT scan from the information contained in a single inspiratory scan at TLC to derive the percentage of lung occupied by functional small airway disease (fSAD^TLC) and by fSAD with emphysema. They applied their generative artificial intelligence (AI) model to a training set of inspiratory CT scans from 1,055 randomly selected participants in the SPIROMICS study. Next, they examined the remaining 1,458 SPIROMICS participants to confirm the relationship between the fSAD data in the single-scan and dual-scan populations. They examined the clinical associations between the fSAD^TLC data and health status, change in lung function at 5 years (in 650 participants), and exacerbation frequency, controlling for multiple relevant confounders including baseline postbronchodilator FEV₁. They extended their approach to 458 COPDGene participants and examined the reproducibility of their method in a subset of 100 SPIROMICS participants with repeat CT scans 2–6 weeks apart (11). Rigorous, if complex, statistical methodologies were applied in examining the similarities between different methods of determining fSAD data.

From this wealth of analysis several clear messages emerge. Using the AI-generated single-scan data produces results that are virtually identical to those obtained from both inspiratory and expiratory scanning. The fSAD measures of small airway disease and localized emphysema independently contribute to a small but measurable increase in the rate of loss of lung function and the chance of experiencing an exacerbation, worse health status, and a higher mortality The magnitude of this effect on FEV₁ decline differed between the SPIROMICS and COPDGene projects, although whether this reflects uncontrolled confounders within the two study populations or slightly differences in scanning protocols is unclear. The fSAD^TLC data were more reproducible than with the dual-scanning approach. This is not surprising, as the variation is inevitably less when two rather than four scans are compared, especially when the supine expiratory scan at residual volume did not need to be repeated.

As the authors note, theirs is not the first study to use AI to demonstrate that a single inspiratory scan can be used to generate data about functional small airway disease (12). However, this is a much larger dataset, with its principal findings confirmed in data from other studies and with good between-tests reproducibility. Unsurprisingly, the clinical data here agree with the earlier findings from dual-scanning studies and illustrate that the clinical problems of patients with COPD and, particularly, disease progression are determined by the totality of damage in the COPD lung and not just by those areas with macroscopic abnormality. Clearly, all the information needed to infer the presence of microscopic lung injury is contained within the inspiratory scan, and it would be interesting to know whether this could be accessed without the need to generate the expiratory virtual scan. Further analyses from COPDGene have shown that the extent of mechanically abnormal lung adjacent to an emphysematous area is a good predictor of future lung function decline, as are ground glass opacities (13, 14). Understanding how this new methodology relates to these other imaging abnormalities will provide further knowledge about how COPD develops in its earliest stages.

Some caution is warranted in interpreting the data of Chaudhary and colleagues. Their findings apply to patients undergoing a specific research scanning protocol, and it would help to compare the results obtained with those in a real-world setting. Nonetheless, it is tempting to speculate whether scans obtained in other research studies such as MESA (15) or ECLIPSE (7) could now be used to extend our knowledge of the importance of fSAD in all stages of COPD.

The most frightening character in the fictional worlds of Arthur C. Clarke was the onboard computer HAL 9000 whose mental instability threatened the lives of those it was there to protect. Undoubtedly, that image has led to caution in our accepting the findings of any machine intelligence, but on this occasion, the ability of AI to offer insights into lung damage that we could not otherwise discern offers real benefits. It seems unlikely that this will be the last time that technology simulates what ordinary people would call magic.