Less experienced radiologists benefit from deep learning models when scouting for intracranial aneurysms
Deep learning models can increase reader accuracy while simultaneously decreasing interpretation times when evaluating imaging for intracranial aneurysms (IA).
Experts came to this definitive conclusion after conducting a meta-analysis that tested the utility of 17 different deep learning models for the detection of intracranial aneurysms. They suggested that their findings could present a solution to some of the issues surrounding IA assessments by addressing the cost-effectiveness (or lack thereof, in this case) and offering diagnostic stability that does not depend on a reader’s experience level.
“The computer-assisted detection program seems to be a solution due to the relative shortage of experienced radiologists compared with the increasing demand for imaging studies,” corresponding author Zhong Wang, from the Department of Neurosurgery & Brain and Nerve Research Laboratory at The First Affiliated Hospital of Soochow University in China, and colleagues discussed. “However, according to our knowledge, there is a lack of relevant meta-analysis regarding the application of deep learning in the clinical diagnosis of IAs.”
Using a mixed-effect binary regression model, the researchers analyzed the performance of the 17 models at patient- and lesion-level, measuring their utility both as stand-alone readers and as assistive tools when assessing CTA, MRA and DSA exams.
At the patient level, the models had high sensitivity and specificity. This finding carried over into sensitivity metrics at the lesion level as well and became most prominent when the models were used as decision support tools for radiologists, achieving the highest sensitivity observed.
The experts noted that historically clinicians have recorded significant variability in their accuracy for detecting IAs, and that performance relies heavily on the reader’s experience level. With this in mind, they suggested that using deep learning models as decision support tools where resources are lacking could be beneficial.
“According to our analysis, the sensitivity of human readers detecting IAs with the assistance of DLMs was significantly higher than those without assistance,” they explained.
The authors acknowledged that although their meta-analysis provided support for incorporating DLMs into clinical workflows, the costs and time required to develop and train them are something that must be considered.
“A cost-benefit assessment of the high cost of the graphical processing unit should be considered. In the coming future, more studies with larger samples and more radiologists should be encouraged.”
The study abstract can be viewed in the European Journal of Radiology.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She joined Innovate Healthcare in 2021 and has since put her unique expertise to use in her editorial role with Health Imaging.