Translational vision science & technology · 2026
AI for Anterior Segment Disease Using Transfer Learning: Adapting Slit-Lamp-Trained Model for Analysis of Smartphone Corneal Images.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Maehara Hiroki, Ueno Yuta, Oda Masahiro, Ito Yoshikazu, Kitaguchi Yoshiyuki, Miyazaki Dai, Nejima Ryohei, Inomata Takenori, Kato Naoko, Chikama Tai-Ichiro, Ominato Jun, Yunoki Tatsuya, Tsubota Kinya, Suzutani Manabu, Mori Kensaku, Yamaguchi Takefumi, Oshika Tetsuro
Аннотация
PURPOSE: The automated classification artificial intelligence (AI) for anterior segment corneal diseases that we developed is capable of classifying images into nine categories, including vision-threatening corneal diseases, such as infectious keratitis; however, it has been trained exclusively on slit-lamp images. We aimed to develop an AI model adaptable to smartphone images by applying transfer learning using smartphone images to the existing AI model. AI trained with transfer learning on smartphone images will be referred to as "Phone-tuned AI."
METHODS: This study included 2530 images captured using smartphones, collected from multiple collaborating facilities between October 2021 and March 2024. Transfer learning was applied to two existing AI models (You Only Look Once version 5 [YOLOv5] and YOLOX) using these smartphone images, and the accuracy of the Phone-tuned AI was evaluated.
RESULTS: The average accuracy for each classification using smartphone images was 93.5% for Phone-tuned AI (YOLOv5) and 67.0% for Original AI (YOLOv5), showing a statistically significant improvement (P = 0.0033). Similarly, the accuracy was 84.2% for Phone-tuned AI (YOLOX) and 78.4% for Original AI (YOLOX), with no significant difference (P = 0.36). When diseases were categorized by urgency, the Phone-tuned AI (YOLOv5) achieved 94.8% for urgent, 89.7% for semi-urgent, 89.4% for routine, and 98.7% for observation-level cases.
CONCLUSIONS: Phone-tuned AI has the potential to assist in diagnosis and triage in regions with a shortage of ophthalmologists, such as rural areas.
TRANSLATIONAL RELEVANCE: Transfer learning using smartphone images showed a particularly good fit with YOLOv5, resulting in high diagnostic accuracy and demonstrating strong potential for clinical application.
METHODS: This study included 2530 images captured using smartphones, collected from multiple collaborating facilities between October 2021 and March 2024. Transfer learning was applied to two existing AI models (You Only Look Once version 5 [YOLOv5] and YOLOX) using these smartphone images, and the accuracy of the Phone-tuned AI was evaluated.
RESULTS: The average accuracy for each classification using smartphone images was 93.5% for Phone-tuned AI (YOLOv5) and 67.0% for Original AI (YOLOv5), showing a statistically significant improvement (P = 0.0033). Similarly, the accuracy was 84.2% for Phone-tuned AI (YOLOX) and 78.4% for Original AI (YOLOX), with no significant difference (P = 0.36). When diseases were categorized by urgency, the Phone-tuned AI (YOLOv5) achieved 94.8% for urgent, 89.7% for semi-urgent, 89.4% for routine, and 98.7% for observation-level cases.
CONCLUSIONS: Phone-tuned AI has the potential to assist in diagnosis and triage in regions with a shortage of ophthalmologists, such as rural areas.
TRANSLATIONAL RELEVANCE: Transfer learning using smartphone images showed a particularly good fit with YOLOv5, resulting in high diagnostic accuracy and demonstrating strong potential for clinical application.