Abstract
Purpose :
To develop an automated classification system for infant fundus image quality based on deep learning, filtering out images of substandard quality with automated feedback on the reasons, addressing the issue of quality variation and facilitating the optimization of Retinopathy of Prematurity (ROP) images.
Methods :
This retrospective study collected RetCam images and clinical data of infants who visited the Shenzhen Eye Hospital from 2004 to 2021. The study was approved by the Ethics Committee of Shenzhen Eye Hospital. It followed the principles of the Helsinki Declaration, and all patient personal information was anonymized before data analysis. All included images were captured by experienced technicians using the RetCam camera after dilating the pupils. A total of 3770 retinal images were included in the development of the AI model. We have developed a multi-level semantic feature aggregation dual-branch classification model that integrates Convolutional Neural Network(CNN) and Transformer. The results are evaluated by area under the curve (AUC), accuracy, precision, recall, and F1 score. We classified images into 4 categories: acceptable, unfocused, overexposed, and underexposed, with each category containing 995, 995, 870, and 870 images respectively.
Results :
This study compared the performance of multiple models based on the ResNet network. After preliminary screening of several base networks, it was found that our network performed the best. It could accurately identify image quality with an average accuracy of 93.79%, precision of 93.74%, recall of 93.87%, F1 score of 93.78%, and an AUC of 97%. The results suggest that the AI model can effectively assess image quality and distinguish various kinds of poor quality images. Further optimization on this network can enhance classification results, providing clear image data for subsequent intelligent diagnostic analysis of various retinal diseases.
Conclusions :
An automatic classification system for the quality of infant retinal images based on deep learning has been developed. This system demonstrates excellent performance in differentiating image quality, promising to provide higher quality images for clinical research and work, thereby enhancing the efficiency of clinical practice.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.