The following is a summary of “Using artificial intelligence to improve human performance: efficient retinal disease detection training with synthetic images,” published in the March 2024 issue of Ophthalmology by Tabuchi et al.
Researchers conducted a retrospective study to introduce a teaching method aided by artificial intelligence (AI). The method utilizes generative AI to train students with numerous images while safeguarding patient privacy.
They developed a web-based course utilizing 600 synthetic ultra-widefield (UWF) retinal images to instruct students in disease detection. Images were generated using stable diffusion, a large generative foundation model fine-tuned with 6,285 real UWF images across six categories, including five retinal diseases (age-related macular degeneration, glaucoma, diabetic retinopathy, retinal detachment, and retinal vein occlusion), along with normal images. The course was completed by 161 trainee orthoptists who underwent evaluation through two tests, one featuring UWF images and another with standard field (SF) images containing 120 real patient images (20 per category). Each student took both tests before and after training, with a cooling-off period in between.
The results showed that the students completed the course in an average of 53 minutes, showing substantial enhancements in diagnostic accuracy. Specifically, for UWF images, student accuracy improved from 43.6% to 74.1% (P<0.0001 by paired t-test), almost reaching the performance level of the previously published state-of-the-art AI model, which achieved an accuracy of 73.3%. In the case of SF images, student accuracy increased from 42.7% to 68.7% (P<0.0001), surpassing the performance of the state-of-the-art AI model, which achieved an accuracy of 40%.
Investigators concluded that while synthetic images enhance medical education, human adaptability surpasses AI, highlighting the irreplaceable role of human judgment in diagnosis.
Source: bjo.bmj.com/content/early/2024/03/14/bjo-2023-324923
