Deploying Artificial Intelligence for Thoracic Imaging Around the World
Samantha Singh, Ameena Elahi, Alan Schweitzer, Ademola Adekanmi, Omolola Atalabi, Daniel J Mollura, Farouk Dako
J Am Coll Radiol . 2023 Sep;20(9):859-862. doi: 10.1016/j.jacr.2023.06.024. Epub 2023 Jul 23.
Purpose: Artificial intelligence (AI) thoracic imaging applications are increasingly being deployed in low- and middle-income countries (LMICs). Radiologists have a critical gatekeeping role to ensure the effective and ethical implementation of AI solutions. RAD-AID International uses a three-pronged implementation strategy to overcome challenges pervasive in LMICs.
Methods: During a similar period, an AI software for chest radiography (CXR) interpretation was deployed at two tertiary hospitals located in Guyana and Nigeria. The three-pronged implementation strategy of clinical education, infrastructure implementation, and phased AI introduction was used. A PACS with a cloud component was installed at each institution. Radiology residents and attending physicians at these institutions completed an introduction-to-AI course to prime them for the use of AI solutions. A phased introduction of the AI software was performed to allow local validation as well as trust building and workflow integration. Local validation processes were used at each site by comparing AI outputs with standardized prospectively generated reports by local radiologists and study team members, allowing for slight differences in the goals of AI software use between sites.
Results: The PACS was successfully installed at both institutions. Thirty participants completed the introduction-to-AI course with an average pre-knowledge test score of 75% and an average posttest score of 95%. The focus of the validation process at various sites was reflective of the intended use of the AI software. In Guyana, it revealed an 87% concordance rate between radiologists and the AI model for triaging normal versus abnormal findings on CXR. In Nigeria, an 85% concordance rate between radiologists and the AI model for reporting tuberculosis on CXR was noted. The AI software was successfully deployed and is being used as intended at both institutions.
Conclusions: There are unique barriers to the adoption of AI in LMICs requiring an implementation strategy in collaboration with local institutions and industry partners to ensure success.