Abstract
Purpose :
This study examined whether diagnostic identification abilities could be improved through training with synthetic images of eyelid tumors generated by an image-creating AI.
Methods :
The diagnostic challenge was distinguishing between chalazion, a representative benign eyelid tumor, and malignant eyelid tumors (sebaceous carcinoma, basal cell carcinoma, squamous cell carcinoma). Using the Stable Diffusion XL model (Stability AI, USA) optimized with Dreambooth (Google and Boston University, USA), we generated 200 synthetic images: 100 of chalazion and 100 of malignant eyelid tumors, each created from photos of real patient anterior eye segments. (Figure 1, Left side: a synthetic image like a chalazion, Right side: a synthetic image like a malignant eyelid tumor ) These images were divided into five sets, and a three-choice learning quiz was created for each set, consisting of 40 images, with 20 images of chalazion and 20 images of malignant tumors, making up 50% of each set. The quizzes were created using Learning BOX (Learning BOX Co., Japan) and were accessible online for answering and explanatory learning. The three choices were "Chalazion," "Malignant Eyelid Tumor," and "Don't Know." To assess learning effects, an evaluation test consisting of 50 real images (25 of chalazion and 25 of malignant tumors) was created and conducted online, similar to the learning quizzes. The training participants were five orthoptists. The total study time was determined from the time taken to access the first question of all 200 until after answering the last question.
Results :
The average pre-study test accuracy of the five participants was 74% (SD 6.9%) (Range 64-80%), which increased to 84% (SD 6.8%) (Range 76 – 92%) post-study, showing a significant improvement (Paired t-test, p=0.0008). The total study time for the five participants averaged 10 minutes and 39 seconds (SD 1 minute and 11 seconds) (Range 9 minutes and 14 seconds – 11 minutes and 18 seconds).
Conclusions :
Synthetic images generated by image-creating AI significantly improved medical image diagnostic identification abilities. Furthermore, the time taken for this study task was short, indicating high efficiency.This demonstrated the potential of the newest generative artificial intelligence technology to shorten the learning curve, which is one of the primary goals of clinical education, as a human-centric technology.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.