Artificial intelligence systems could be deployed to evaluate CT scans in after hours hospital periods when there is a shortage of human radiologists, Australian researchers have suggested.
Radiology reporting based on deep learning algorithms has the potential to improve diagnostic accuracy, reduce reading times and prioritise abnormal findings, which may be valuable benefits in the after-hours period when senior radiology consultants may not be available, according to a new paper.
Authored by radiologists at Royal North Shore Hospital in Sydney, the study found inaccuracies in more than one tenth of after-hours CT scan reports their own trainees had created in 2019. The analysis of almost 11,000 reports covered conditions such as suspected cervical spinal fractures, stroke investigation, CT brain intracranial haemorrhage detection, pulmonary embolism detection and rib fracture.
Some 3.4% contained “significant errors”, defined as errors impacting patient management or potentially accounting for the patient’s clinical presentation.
These included a missed case of reversible cerebral vasoconstriction syndrome on CT, as well as a missed three column T8 spinal fracture.
With the hospital’s radiology department staffed exclusively by registrars after 4 pm, the authors stressed the error rates were not unusual compared with other studies into trainee reporting accuracy.
All inaccuracies were identified by a consultant radiologist the following business day, they added in the Journal of Medical Imaging and Radiation Oncology.
They pointed out error rates fell with experience, with 16% of reports created by first-year trainees including at least one mistake and 18% of those written by second-years.
By contrast, doctors in their later years of training had issues in their reports just 11% of the time.
Nevertheless, this raised a question as to why artificial intelligence was not being used, particularly given the demands of busy after-hours shifts, the authors said.
“In the emergency radiology setting, AI has the potential to have a positive clinical impact by improving diagnostic accuracy, reducing reading times and prioritising abnormal studies,” the authors wrote.
“This may assist after-hours hospital reporting in the future and help manage increasing caseloads.”
The issue was particularly timely given there were now at least 100 commercially AI radiology products on the market, all based on deep learning algorithms, they said.
Available emergency products included automated large vessel occlusion detection in CT angiography for stroke investigation, CT brain intracranial haemorrhage detection, pulmonary embolism detection, cervical spine and rib fracture detection on CT.
“An example of practical use is improved sensitivity of intracranial haemorrhage detection and assistance prioritising abnormal cases for reporting to expediate clinically significant diagnoses,” they added.