Most of the currently published deep-learning studies in medical image analysis report performance using carefully selected datasets. It's critical, though, to know how the algorithms work in the real world, according to presenter Hyunkwang Lee, PhD, of Massachusetts General Hospital.
The researchers applied their deep learning-based intracranial hemorrhage detection system to all noncontrast head CT scans acquired at an emergency department over a three-month period. They then compared the results with the algorithm's performance on selected cases from the dataset that was initially used for training and validating the model.
While the algorithm yielded similar negative predictive value on the real-world images, it also produced lower sensitivity and specificity, according to the researchers.
"This performance gap could be attributable to exclusion of cases with any history of brain surgery, intracranial tumor, intracranial device placement, skull fracture, or cerebral infarct during model development and the presence of CT artifacts incurred by using different imaging systems with varying scan acquisition and reconstruction settings," Lee told AuntMinnie.com. "By understanding the nature of incorrect prediction in the real-world data, the model can be improved through constant feedback from expert radiologists, facilitating the adoption of such tools in clinical practice."
What else will the researchers report? Attend the Artificial Intelligence in Neuroradiology refresher course on Tuesday morning to find out.
