Share on Facebook
Share on Twitter
Share on LinkedIn
Share by Email
Print
Among the rapid advances in the use of artificial intelligence (AI) in medicine in recent years, a proliferation of studies demonstrates its value specifically in pulmonology. Such technology has the potential to ease the workload of physicians while increasing the speed and accuracy of diagnosis, and a growing body of research highlights the role of AI in diagnostic and prognostic imaging for pulmonary diseases.
In lung cancer screening, AI models have shown the ability to automatically identify pulmonary nodules and differentiate between benign and malignant nodules.1 In addition, a range of “…studies in non-small cell lung cancer (NSCLC) used radiomics to predict distant metastasis in lung adenocarcinoma and tumour histological subtypes as well as disease recurrence, somatic mutations, gene-expression profiles and overall survival,” according to a 2018 review published in Nature Reviews Cancer.1
Among the more recent observations in this area, in 2019, researchers developed and tested a deep-learning based algorithm in classifying results of chest radiographs (the most widely performed radiographic test, encompassing up to 26% of diagnostic imaging exams currently) from patients with pulmonary malignant neoplasm, active tuberculosis, pneumonia, and pneumothorax. The model was developed using 54,221 radiographs with normal findings and 35,613 with abnormal findings.2
The algorithm showed consistently higher performance than non-radiology physicians, board-certified radiologists, and thoracic radiologists in both image-wise classification (area under the receiver operating characteristic curve [AUROC], 0.983 vs 0.814, 0.896, and 0.932, respectively; all P <.005) and lesion-wise localization (area under the alternative free-response receiver operating characteristic curve [AUAFROC], 0.985 vs 0.781, 0.870, and0.907, respectively; all P <.001).
These and related study results highlight the potential for AI to “… help radiologists identify and prioritize patients with the most severe or complex cases” and demonstrated its “… effectiveness as a second reader in conjunction with a radiologist by, for example, enhancing lesion localization,” as stated in a paper published in January 2020 in the Journal of the American College of Radiology.3
However, there remain various challenges for researchers to address before the full value of AI can be realized in diagnostic imaging for pulmonary and other diseases. For example, AI models may be less useful in certain diseases that share manifestations with other diseases.
In a study published online in July 2020 in the American Journal of Roentgenology, AI software did not detect significant differences between COVID-19 pneumonia and influenza virus pneumonia on computed tomography (CT) because of the large degree of overlap in CT manifestations of both conditions.4 This finding also underscores the limited utility of relying on imaging or AI alone in distinguishing between diseases.
For additional insights regarding progress and remaining issues in this field, we interviewed Luciano Prevedello, MD, MPH, associate professor of radiology, vice-chair for medical informatics and augmented intelligence in imaging, division chief of medical imaging informatics, and medical director of the 3D and Advanced Visualization Lab at The Ohio State University Wexner Medical Center in Columbus; and Eduardo J. Mortani Barbosa, Jr, MD, assistant professor of radiology at the University of Pennsylvania’s Perelman School of Medicine in Philadelphia.
What have been some of the most notable recent developments in the use of AI in diagnostic imaging — for pulmonology specifically, and what is the state of this field currently?
Dr Prevedello: The amount of research and innovation in this field is quite impressive. Deep learning has revolutionized the field of computer vision and has the potential to make a significant positive effect in medical imaging. The technology is quite capable in assisting radiologists with lesion detection and characterization (such as detection of pulmonary nodules and differentiation between benign and malignant lesions), image segmentation (such as segmentation of the lungs or ground-glass areas of lung abnormality in patients with COVID-19), and improving both speed and quality of image acquisition.4
The majority of discoveries in this area are still in the research setting. It takes time to transfer these technologies into the clinical setting because they need to be thoroughly tested and evaluated.
Dr Barbosa: There have been numerous published AI research articles applied to diagnostic imaging, noting exponential growth in the past 2 decades. Therefore, I will focus on a broad, high-level overview of pulmonary imaging applications. Conceptually, AI applied to pulmonary imaging (CT, magnetic resonance, chest x-ray) can accomplish one or more of these tasks: disease detection, disease segmentation/quantification, and disease classification.
When combined with patient-centered outcomes, AI can also drive prognostic prediction models. In terms of specific pulmonary applications, 2 large fields include risk assessment for malignancy in lung nodules and quantification of diffuse lung diseases such as chronic obstructive pulmonary disease and interstitial lung disease. For some very constrained and well-defined tasks, the best AI systems can perform as well as experts.
What are some of the remaining challenges and research needs in this area?
Dr Prevedello: One important current limitation of the technology is the issue of generalizability/transferability of the models. Algorithms created with data from one single hospital may not perform well in another hospital or on scanners with different parameters. One way to address the issue is by including a very representative and heterogeneous set of images when training these algorithms. Having data from multiple organizations is not easy because of patient privacy concerns, but there is also a lot of research going on in this area to ensure this can be done in a safe way. For example, federated learning is a method in which data stays within the organization and only the model weights are shared, therefore no patient-specific data is shared.1
Dr Barbosa: There are many challenges, including data quality and research design: Without high-quality, curated data with best possible ground truth, and without meaningful, well-designed research questions, results can be of dubious validity. Another issue is data diversity — if it is limited, this will reduce the generalizability of the AI models. The third challenge is the inherent complexity of lung diseases and frequent overlapping imaging patterns (for example, many distinct diseases presenting as a similar imaging pattern), as well as coexistence of distinct diagnoses leading to a multitude of patterns in the same patient at the same time point.
Therefore, it is unlikely that AI systems that rely solely on imaging will replace the integrative reasoning and judgment of an astute physician, at least in the foreseeable future. The integration of knowledge and different sources of information, combined with longitudinal assessment of patient data over time, will remain crucial for accurate diagnosis and optimal patient management.
What are the current clinical implications regarding the use of AI in diagnostic imaging?
Dr Prevedello: The field is evolving quite rapidly, and many applications are receiving US Food and Drug Administration approval and making their way into the clinical environment. These tools are showing promising results and will likely have an important role in assisting us to continue to provide better care to our patients.
Dr Barbosa: Do not rely blindly on published results without a critical appraisal of the quality of the data and study design, as well as validity and generalizability of the results. AI is a powerful and disruptive technology which will change the way we practice medicine and will likely lead to improvements in diagnosis and prognosis at the population level once widely implemented. However, this will happen as an augmentation of physician-driven care, rather than as a replacement.
References
1. Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nat Rev Cancer. 2018;18(8):500-510.
2. Hwang EJ, Park S, Jin KN, et al. Development and validation of a deep learning-based automated detection algorithm for major thoracic diseases on chest radiographs. JAMA Netw Open. 2019;2(3):e191095. doi:10.1001/jamanetworkopen.2019.1095
3. Alexander A, Jiang A, Ferreira C, Zurkiya D. An intelligent future for medical imaging: a market outlook on artificial intelligence for medical imaging. J Am Coll Radiol. 2020;17(1 Pt B):165-170.
4. Lin L, Fu G, Chen S, et al. CT Manifestations of coronavirus disease (COVID-19) pneumonia and influenza virus pneumonia: a comparative study. AJR Am J Roentgenol. Published online July 9, 2020. doi:10.2214/AJR.20.23304
