Screening for chronic obstructive pulmonary disease with artificial intelligence

Abstract

The use of artificial intelligence (AI) to perform medical tasks is being assessed in almost every field of practice. Deep neural networks, which were first used to classify pictures of objects, are now being repurposed to diagnose skin cancer,1Esteva A Kuprel B Novoa RA et al.Dermatologist-level classification of skin cancer with deep neural networks.Nature. 2017; 542: 115-118Crossref PubMed Scopus (54) Google Scholar screen for breast cancer,2McKinney SM Sieniek M Godbole V et al.International evaluation of an AI system for breast cancer screening.Nature. 2020; 577: 89-94Crossref PubMed Scopus (856) Google Scholar predict circulatory failure,3Hyland SL Faltys M Hüser M et al.Early prediction of circulatory failure in the intensive care unit using machine learning.Nat Med. 2020; 26: 364-373Crossref PubMed Scopus (104) Google Scholar and perform image reconstruction,4Shen L Zhao W Xing L Patient-specific reconstruction of volumetric computed tomography images from a single projection view via deep learning.Nat Biomed Eng. 2019; 3: 880-888Crossref PubMed Scopus (103) Google Scholar among many other applications. In The Lancet Digital Health, Lisa Y W Tang and colleagues5Tang LYW Coxson HO Lam S Leipsic J Tam RC Sin DD Towards large-scale case-finding: training and validation of residual networks for detection of chronic obstructive pulmonary disease using low-dose CT.Lancet Digital Health. 2020; (published online April 21.)https://doi.org/10.1016/S2589-7500(20)30064-9Summary Full Text Full Text PDF Scopus (29) Google Scholar report the use of a transfer learning approach to detect signs of chronic obstructive pulmonary disease (COPD) on low-dose CT. COPD is a life-threatening lung disease with a global prevalence of 250 million. Each year, COPD disease accounts for 5% of all deaths in the world.6WHOChronic obstructive pulmonary disease (COPD).https://www.who.int/news-room/fact-sheets/detail/chronic-obstructive-pulmonary-disease-(copd)Date: Dec 1, 2017Date accessed: March 12, 2020Google Scholar COPD is likely to increase in coming years due to higher smoking prevalence, environmental changes, and aging populations in many countries.6WHOChronic obstructive pulmonary disease (COPD).https://www.who.int/news-room/fact-sheets/detail/chronic-obstructive-pulmonary-disease-(copd)Date: Dec 1, 2017Date accessed: March 12, 2020Google Scholar Although the disease is not curable, medical and physical treatments can help to relieve symptoms and improve exercise capacity and quality of life. Diagnosis of COPD is done by spirometry; however, it is considered to be considerably underdiagnosed. Population studies have assessed that 10% of adults aged 40 years or older have airflow limitation on spirometry, but only 20–30% of these individuals have been diagnosed with COPD, suggesting that approximately 70% of individuals with COPD worldwide may be undiagnosed.7Diab N Gershon AS Sin DD et al.Underdiagnosis and overdiagnosis of chronic obstructive pulmonary disease.Am J Respir Crit Care Med. 2018; 198: 1130-1139Crossref PubMed Scopus (121) Google Scholar In this context, the method of repurposing low-dose CT (usually performed for lung cancer screening) for the diagnosis of COPD, as reported by Tang and colleagues, is interesting because of an epidemiological association between these two diseases: the chronic inflammation that plays a role in the pathogenesis of lung cancer is also associated with COPD. Inflammation in COPD results in repeated airway epithelial injury, high cell turnover rates, and an increase in DNA errors, resulting in the amplification of the carcinogenic effects of cigarette smoke.8Durham AL Adcock IM The relationship between COPD and lung cancer.Lung Cancer. 2015; 90: 121-127Summary Full Text Full Text PDF PubMed Scopus (223) Google Scholar Therefore, patients getting screened for lung cancer could simultaneously be screened for COPD with a single modality, for a potential low cost, and without increasing the time required for physician analysis. To our knowledge, the use of deep learning for COPD detection has been assessed in only one other study by González and colleagues,9González G Ash SY Vegas-Sánchez-Ferrero G et al.Disease staging and prognosis in smokers using deep learning in chest computed tomography.Am J Respir Crit Care Med. 2018; 197: 193-203Crossref PubMed Scopus (142) Google Scholar in which the network was trained on a larger cohort of 9300 labelled images.9González G Ash SY Vegas-Sánchez-Ferrero G et al.Disease staging and prognosis in smokers using deep learning in chest computed tomography.Am J Respir Crit Care Med. 2018; 197: 193-203Crossref PubMed Scopus (142) Google Scholar The appendix of the report by Tang and colleagues includes a comparison of their study with that of González and colleagues. A major difference between the two studies is that Tang and colleagues used an external test dataset to verify the model's generalisability: in the ECLIPSE cohort, the models achieved an AUC of 0·886 (SD 0·017), with a positive predictive value of 0·847 (0·056) and a negative predictive value of 0·755 (0·097). The authors should also be commended for providing other evaluation metrics that are more appropriate to report, considering the class imbalance of the ECLIPSE dataset, in which 77·9% of the participants had COPD. These metrics (F1-score, precision, and recall) also show that the models provide robust performance. The article by Tang and colleagues also has some limitations. Beyond the detection of COPD, as a simple binary endpoint, the authors provide activation maps that could be used to assess interpretability. These maps show the relative importance of each pixel in contributing to the classification result. The authors report that the model observed edges of the trachea, some parts of the ribs, vessels, and bronchial walls, which is not in concordance with COPD being a disease of the lung itself. This surprising result could be explained by an association of COPD with tracheal or pleural inflammation. Ablation tests could be performed to better understand this result. The model could also be basing some of its prediction on image artifacts, as is shown in one of the example CT scans in the appendix. A similar phenomenon has been described in the context of chest x-rays.10Zech JR Badgeley MA Liu M Costa AB Titano JJ Oermann EK Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study.PLoS Med. 2018; 15e1002683Crossref PubMed Scopus (508) Google Scholar Finally, we should also consider how actionable the results of this model are. Provided we use such an approach to detect COPD in undiagnosed patients during lung cancer screening, can we define strategies to effectively treat these patients? Will an automatic detection require a systematic confirmation on spirometry, and would we be able to translate that into better care? These questions are the next step in developing medical AI systems: AI is a potentially powerful tool, but can we leverage it for better medicine? We declare no competing interests. Towards large-scale case-finding: training and validation of residual networks for detection of chronic obstructive pulmonary disease using low-dose CTOur proposed approach could identify patients with COPD among ex-smokers and current smokers without a previous diagnosis of COPD, with clinically acceptable performance. The use of deep residual networks on chest CT scans could be an effective case-finding tool for COPD detection and diagnosis, particularly in ex-smokers and current smokers who are being screened for lung cancer. Full-Text PDF Open Access

Topics & Concepts