Radiology Implementation Considerations for Artificial Intelligence (AI) Applied to COVID-19, From the AJR Special Series on AI Applications

Abstract

The publication of this Accepted Manuscript is provided to give early visibility to the contents of the article, which will undergo additional copyediting, typesetting, and review before it is published in its final form. During the production process, errors may be discovered that could affect the content of the Accepted Manuscript. All legal disclaimers that apply to the journal pertain. The reader is cautioned to consult the definitive version of record before relying on the contents of this document.

Data standards and standardization: The shortest plank of bucket for the COVID-19 containment

“…The major reasons for such challenge include inconsistency in data standards and standardization, lack of usability for laypersons, difficulty of deployment in resource-poor settings, and potential ethical pitfalls or legal barriers. 20 The systems with the highest success rate of migration were the classification of chest CT images based on artificial intelligence (AI) technologies 21 since the data in the Picture Archiving and Communication System (PACS) around the world follow the Digital Imaging and Communication in Medicine (DICOM) standard. However, the power of AI and datadriven predictive science played little role in improving the general level of clinical care for the COVID-19 patients, especially for the severe cases as the data infrastructure of standards and standardization were not ready for such challenges.…”

AI-Supported Comprehensive Detection and Quantification of Biomarkers of Subclinical Widespread Diseases at Chest CT for Preventive Medicine

“…Studies have shown the advanced abilities of automated artificial intelligence (AI)-based imaging analysis in Radiology, especially for two reasons: complex image analyses can be performed user-independently within second-short timeframes and were shown to perform up to 11.5% percent better than radiologists, e.g., in breast cancer prediction with an AUC of 0.889 (95% CI 0.871, 0.907; n = 25,856 patients) [ 21 , 22 , 23 , 24 ]. Similar to increasing sub-specialization in medicine, available computer-based image analysis tools are focused on specified tasks, e.g., typically “emphysema”, “pulmonary embolism”, “nodule”, “aortic dissection” or more recently, “COVID-19 pneumonia” giving a precise answer to each question in disregard of coexisting pathologies [ 25 , 26 , 27 , 28 , 29 ].…”

Developing medical imaging AI for emerging infectious diseases

“…The COVID-19 pandemic arose during one of the most innovative periods for biomedical data science, chiefly led by achievements in artificial intelligence (AI), which inspired many efforts globally to leverage digital health data and machine learning techniques to address challenges posed by the pandemic. However, of the innumerable COVID-19-related machine learning efforts developed during the pandemic's most critical time, almost all failed to materialize demonstrable value, and worse, some were potentially harmful [1][2][3][4][5] . That these failures occurred despite the AI and data science community's worldwide united attempt to generate, aggregate, share and utilize large volumes of COVID-19-related data, ranging from simple dashboards, to models for populational-wide risk prediction 6,7 , early detection and prognostication [8][9][10][11][12] , severity scoring [13][14][15] , long-term outcome and mortality predictions [16][17][18][19] , is alarming.…”