"A High-quality Large-scale Color Fundus Photography Images Dataset with Eleven Types of Annotations"

"More and more human beings are suffering from ophthalmic diseases, and automatic analysis of images from Color Fundus Photography (CFP) using deep learning algorithms can improve the accuracy and prevalence of early screening for ophthalmic diseases. However, the performance of deep learning models is highly dependent on the training data, and the currently available retinal image datasets still have limitations such as scarce total data and poor disease diversity. Therefore, we extensively collected and integrated the CFP publicly available datasets, resulting in the construction of the LCFP-14M dataset containing 13,718,610 retinal fundus images with a total of 150,904,710 annotations, which is the largest dataset and covers the widest range of ophthalmic diseases at present. In the experimental validation, we use the LCFP-14M dataset for model fine-tuning and pre-training respectively. The results demonstrate the reliability of the image quality of the LCFP-14M dataset, and the base model obtained by pre-training with LCFP-14M has stronger generalization and convergence speed, which shows its great potential in the field of general ophthalmic disease diagnosis."