Abstract
Purpose :
This study addresses the critical need for microaneurysm segmentation in fundus fluorescein angiography (FFA) images, an essential tool in the diagnosis and management of diabetic retinopathy (DR). We present the development and evaluation of an automated approach, leveraging deep learning, to achieve microaneurysm segmentation in FFA images.
Methods :
A dataset comprising 263 FFA images of 7 patients. Out of this dataset, 100 images underwent manual annotation by three trained graders to identify microaneurysms. For model training, a subset of 75 images was randomly selected, while the remaining 25 were reserved for testing purposes. Two distinct deep neural networks were trained for segmentation, including a coarse network designed to identify microaneurysms easily discernible by the human eye, and a fine network tailored to segment microaneurysms that pose a challenge for human visual detection. Both networks share a UNet architecture with a ResNet backbone. The coarse model underwent supervised training, while the fine model was trained in a weakly-supervised manner. The final segmentation results were derived by overlaying the outputs of these two networks.
Results :
The segmentation of microaneurysms in FFA images was notably accurate using our deep learning-based approach. Specifically, in our testing dataset, we attained average precision, recall, and F1 score values of 0.813, 0.695, and 0.711, respectively, underscoring the robust performance and reliability of our model.
Conclusions :
Our deep learning-based approach of the automated segmentation of micro-aneurysms in FFA images shows encouraging results for enhancing the diagnosis and management of DR. Moreover, our method has the potential to streamline the process of delivering prompt and accurate diagnoses, thereby saving both patients and healthcare providers substantial time and resources.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.