Data Utilization
This is a model that estimates the age from fundus images, created using data from health checkup facilities collected by the Japan Ocular Imaging Registry. While its performance has been validated in internal verification, it does not guarantee its performance with external data. The model is intended to be used in ophthalmology research that utilizes fundus images.
The model is available for free use, and the Japan Ophthalmological Society, Japanese Society of Artificial Intelligence in Ophthalmology, Japan Ocular Imaging Registry, and National Institute of Informatics do not claim ownership of any outcomes resulting from the use of this model. The Japan Ophthalmological Society, Japanese Society of Artificial Intelligence in Ophthalmology, Japan Ocular Imaging Registry, and National Institute of Informatics are not responsible for any output or other results generated by the model.
If you use this model, please cite the following paper and acknowledge the Japan Ophthalmological Society, Japan Ocular Imaging Registry, and National Institute of Informatics in your acknowledgments.
[Citation]
Miyake M, Akiyama M, Kashiwagi K, Sakamoto T, Oshika T. Japan Ocular Imaging Registry: a national ophthalmology real-world database. Jpn J Ophthalmol. 2022 Nov;66(6):499-503. doi: 10.1007/s10384-022-00941-0. Epub 2022 Sep 23. PMID: 36138192.
[Acknowledgment Example]
"The pre-training model used in this study was developed by the Japanese Ophthalmological Society and the National Institute of Informatics, and was made available through the Japan Ocular Imaging Registry website（http://www.joir.jp/)."
Retinal images and health examination data were collected from a single health examination facility. The dataset consisted of 12,734 images from 12,734 individuals (7,375 right eyes and 5,359 left eyes) who had no abnormalities in their retinas and were “healthy”*. The left eye images were flipped horizontally and used for training as if they were all right eye images.
* "healthy" was defined as meeting the following conditions:
Out of 12,734 retinal images, we used 8,149 images for training, 2,038 for validation, and 2,547 for testing. The actual age of the training data was 42.2 ± 11.3 years (minimum 18 years, maximum 84 years) and the distribution is shown in Figure 1. We created an AI model by transfer learning using the Swin Transformer model1 trained on Imagenet2, with the PyTorch library3 in Python34. The input image size was set to 384×384, and an example of an input image is shown in Figure 2.
The actual age of the test data was 42.3 ± 11.0 years. The scatter plot of actual and predicted ages is shown in Figure 3. The mean absolute error (MAE) between actual age and age predicted by the AI model was 2.39 years.
The AI model has been made available in the pth format, and a sample of its operation has been described separately in the source code. Python3 and Pytorch environments are required for operation.
