Related Experiment Video
Updated: Jun 30, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
Using artificial intelligence to improve human performance: efficient retinal disease detection training with
Hitoshi Tabuchi1,2, Justin Engelmann3, Fumiatsu Maeda4
1Department of Technology and Design Thinking for Medicine, Graduate School of Biomedical and Health Sciences, Hiroshima University, Hiroshima, Hiroshima, Japan htabuchi@hiroshima-u.ac.jp.
The British Journal of Ophthalmology
|March 14, 2024
Summary
Generative AI created synthetic retinal images for training, significantly improving diagnostic accuracy in medical students. This AI-aided teaching method highlights the irreplaceable role of human judgment in diagnostics.
Area of Science:
- Ophthalmology
- Medical Education
- Artificial Intelligence
Background:
- Artificial intelligence (AI) offers significant potential in medical imaging diagnostics, yet human judgment remains crucial.
- A novel AI-aided teaching method is proposed, utilizing generative AI to train students on diverse medical images while ensuring patient privacy.
Purpose of the Study:
- To evaluate the efficacy of a web-based course using synthetic ultra-widefield (UWF) retinal images for training medical students in disease detection.
- To assess the impact of AI-generated synthetic images on diagnostic accuracy compared to traditional methods and AI models.
Main Methods:
- A web-based course was developed featuring 600 synthetic UWF retinal images generated by fine-tuned stable diffusion models.
- The training dataset comprised 6285 real UWF images across six categories, including five retinal diseases and normal cases.
- 161 trainee orthoptists completed the course and were assessed using pre- and post-training tests with both UWF and standard field (SF) real patient images.
Main Results:
- Students demonstrated a significant improvement in diagnostic accuracy after completing the course, with UWF image accuracy rising from 43.6% to 74.1% (p<0.0001).
- Diagnostic accuracy for SF images also increased substantially, from 42.7% to 68.7% (p<0.0001), surpassing a state-of-the-art AI model's performance.
- The AI-aided training nearly matched the accuracy of a leading AI model for UWF images and exceeded it for SF images.
Conclusions:
- Synthetic medical images generated by AI are effective tools for medical education and training.
- Human diagnostic performance proved more robust in novel situations than AI, underscoring the continued essential role of human judgment in medical diagnosis.
