Radiology: Artificial Intelligence
This paper explores the applicability of artificial intelligence (AI) in the clinical practice. AI would be deemed appropriate for application if it has accurate diagnostic abilities. In the cases of intracranial hemorrhages, AI shows potential to be a reliable system to aid physicians in diagnosing patients in a timely manner. Intracranial hemorrhage (ICH) is a neurological emergency that is time-sensitive, as brain tissue starts dying within the first hour of onset. The ICH is associated with the mortality rate of 44% at 30 days and the survivors also suffer from severe disability. Non-contrast CT is the gold standard imaging modality for detecting ICH. AI has yet to be implemented in a standardized framework to automatically outline intracranial hemorrhages, but it could greatly speed up the investigational process. Method: All emergent noncontrast CT images (256 sections) were obtained between Nov 1, 2018 and Oct 31, 2020 at University Hospital of Basel (Switzerland). Their reports were also received with approval by board-certified neuroradiologist. AI Algorithm is provided by the deep learning algorithm (AIdoc Medical) was used. If analysis yields positive results, it prompts an alert highlighting the image, but did not reprioritize the worklist. Diagnostic performance was evaluated by calculating diagnostic accuracy, sensitivity, specificity, and predictive values. Result: 4450 patient head CT were evaluated, and 3017 of them were done after AI was implemented. The degree of diagnostic accuracy was as follows:
The AI showed high accuracy at detecting intraventricular hemorrhages (100%), but not so much for subarachnoid hemorrhage or subdural hemorrhage (70-80%). With a degree of low specificity, the AI may give a false sense of security. Implementation of the AI also expedited the time it took to rule out ICH – from 205 minutes to 167 minutes, and thus shortened the turnaround time in the emergency department. The study observes that the diverse practices, false alarms, limitations of human-machine interoperability and more factors in different scenarios may have acted against the improved time efficiency, and emphasizes this as a reason for further integration of AI in the workplace and defining AI frameworks.
References