Radiomics and Computerized Analysis of CT Images: Looking Forward

See also article by Chassagnon et al in this issue.

Brett M. Elicker, MD, is a clinical professor in the department of radiology and biomedical imaging at the University of California, San Francisco (UCSF). He did a radiology residency at Yale and thoracic imaging fellowship at UCSF. His clinical and research interests are in the areas of diffuse lung disease and lung cancer.

Jae Ho Sohn, MD, is a radiologist and researcher at the UCSF Center for Intelligent Imaging (ci²). His research focuses on computer vision for data-driven imaging biomarker discovery in cardiothoracic imaging and natural language processing in radiology.

A radiologist’s interpretation of CT findings often focuses on the detection of abnormalities and the formulation of a differential diagnosis; however, there are other roles that may be equally important. One of these is the ability to correlate CT findings with the risk of disease progression and mortality. While visual assessment still forms the cornerstone of the formulation of a radiology report, there is a potential for these roles to be supplemented or supplanted by computerized analyses of the CT data. This may be particularly pertinent for diseases that are characterized by complex structural abnormalities, such as those seen in interstitial diseases such as idiopathic pulmonary fibrosis or airway diseases such as cystic fibrosis. Visual interpretation is time-consuming and subject to significant interobserver variability, whereas well-designed computerized analysis potentially provides useful information in a rapid fashion and without the downside of poor interobserver agreement.

Cystic fibrosis is a common genetic disorder affecting approximately 1 in 3200 live births in White individuals (1). It is characterized by progressive airway inflammation and dilatation that may ultimately require lung transplantation or lead to death. CT is commonly performed in patients with cystic fibrosis to assess for the presence and nature of morphologic abnormalities. While pulmonary function tests and, more specifically, forced expiratory volume in 1 second (FEV₁), are often used as the primary metrics to evaluate the severity of disease, FEV₁ has significant limitations, including poor reproducibility and less than ideal correlation with the extent of disease. Given the ability of CT to characterize and quantify airway abnormalities in cystic fibrosis, it could provide a superior alternative to FEV₁ in determining patients’ prognoses and might be able to identify patients at risk for more rapid progression of their disease. This information could potentially be used to prospectively identify patients in need of more aggressive treatment or who might be appropriate candidates for clinical trials. Additionally, it could also be used to identify patients who might require lung transplantation in the near future.

Radiomics refers to a technique in which a large number of imaging features are extracted and characterized using computer algorithms. A variety of different types of features may be analyzed, from simple to complex. Arguably, the simplest involves an analysis of the density of pixels (histogram analysis) including features such as mean density, kurtosis (the weighting of the tail of the histogram distribution compared with normal), and skewness (a measurement of the asymmetry of histogram distribution compared with normal), among others. More complex techniques include texture analysis, which studies spatial heterogeneity of the pixels, and fractal analysis, which studies higher-order complex spatial relationships. Recently, deep learning has also become a popular approach to enhancing feature extraction and analysis when working with large amounts of data. Radiomics is a field still in search of clinical usefulness; however, it shows promise in potentially identifying specific phenotypes of diseases that may be associated with prognosis or the risk of future disease progression.

In this issue of Radiology: Cardiothoracic Imaging, Chassagnon et al (2) investigate the use of radiomics in patients with cystic fibrosis. A total of 38 different CT features were processed through five different machine learning techniques in hope of developing models that had reasonable correlation with the risk of disease progression, prognosis, and other metrics of disease severity. Given the relatively poor correlation between pulmonary function test abnormalities and disease severity in cystic fibrosis, a variety of cystic fibrosis clinical scoring systems have previously been developed. Several examples of these scoring systems are the Nkam (3), Liou (4), and CF-ABLE (5) scores. These systems integrate a combination of factors including, but not limited to, the following: age, sex, FEV₁, body mass index, number of intravenous antibiotic courses, need for hospitalization, need for oral corticosteroids, need for long-term oxygen therapy, and need for noninvasive ventilation. These clinical scoring systems have shown to have good correlation with prognosis and mortality. In this study, the CT scoring algorithms were predominantly correlated with these clinical scoring systems. The five CT models that were developed demonstrated moderate to strong correlations with the Nkam score (correlation coefficient [R] = 0.57 to 0.63). Moderate to strong correlations were also seen between the models and the Liou score (R = −0.55 to −0.65), FEV₁ (R = −0.62 to −0.66), and the risk of exacerbations in the 12-month follow-up period (R = 0.38 to 0.55). The derivation and test cohorts for this study were performed at different institutions, suggesting that these models may be generalizable to a variety of different CT scanners with different protocols. Specifically, in this study one significant difference between the two institutions was that the CT scans used to develop the algorithms were reconstructed with filtered back projection, whereas the CT scans in the test cohort were predominantly reconstructed with iterative reconstruction.

The ability to leverage computerized analysis of CT data has, in the past, yielded several potential roles that may provide supplemental information in a radiology report or, in some cases, may provide a more accurate analysis than a radiologist’s visual interpretation. Attenuation analysis has been studied most extensively, specifically as a method of quantifying emphysema (6). By identifying pixels below a threshold Hounsfield unit measurement, the severity of emphysema can be quantified, and progression of disease may be assessed over time. Threshold pixel analyses have also been performed on expiratory images to identify patients with airways-predominant chronic obstructive pulmonary disease. More complex textural algorithms, including deep learning algorithms, have become useful in the detection and risk stratification of pulmonary nodules (7). Complex textural analysis has been developed for use in patients with interstitial lung disease to identify specific CT findings such as reticulation, traction bronchiectasis, and honeycombing. The extent of these abnormalities using textural analysis has been shown to correlate with prognosis and may be more sensitive to changes over time than visual assessment (8).

The current study should be viewed within the context of this larger body of work which has provided insights into diseases that were not apparent using visual analysis alone. While computerized analysis of CT images has found a few clinical indications, many of the algorithms are still in search of clinical utility, in part awaiting rigorous confirmation of generalizability and model calibration. As the algorithms continue to improve, their clinical relevance will undoubtedly also follow and allow for supplementing the information provided by visual analysis of CT alone. Deep learning in particular may allow for the development of even more accurate algorithms that maximize the ability of computerized analysis to detect, diagnose, and characterize disease. The current study, which leverages radiomics, represents a foundation upon which future analyses of patients with cystic fibrosis may be based. With the development of even more accurate analyses and further external validation, CT could potentially supplant other metrics in the identification of patients at risk for rapid progression and death. Radiologists should think beyond detection and diagnosis as the primary clinical uses of CT and should embrace other indications that may be able to identify unique phenotypes within a single disease. While computerized analysis is unlikely to replace radiologists any time soon, it is incumbent upon us to understand the additional information that these analyses provide, the unique strengths and potential pitfalls of these models, and how the clinical management of patients may be improved by this information.