In this research CT images possibly containing intracranial hemorrhage are systematically ranked by exploiting the varying performance of 70 trained deep neural networks. Using the exploited information, a computer-assisted learning (CAL) system is developed, aimed at improving intracranial hemorrhage diagnostic skills in radiology trainees.
By exploiting the varying performance of the trained deep neural networks, possible intracranial hemorrhage cases in a test dataset are assigned a rank using item analysis. Using item analysis classification guidelines found in literature, the cases are respectively put into a level system. An application containing the CAL-system is developed in MATLAB, in which this level system is incorporated. To evaluate the effectiveness of the CAL-system between different groups, participants who respectively have either a medical background or a non-medical background have been recruited for this research to obtain user performance data, who are given a test before and after the CAL-training.
The accuracy of the participants with a non-medical background increased from 69.6% to 79.2% after CAL-training, showing a significant increase in diagnostic performance (p = 0.0125). However, the accuracy of the participants with a medical background stayed the same at 81.0% after CAL-training, showing no significant increase in diagnostic performance (p > 0.05).
Overall, the findings in this research suggest that the CAL-system developed in this research has potential for the training of freshmen radiology students and/or people with no medical background, to bring their intracranial hemorrhage diagnostic skills to a baseline level.
To join the presentation via Microsoft Teams click here