Abstract
Vision loss caused by diabetic macular edema (DME) can be prevented by early detection and laser photocoagulation. As there is no comprehensive detection technique to recognize non-perfusion areas (NPA), we proposed an automatic detection method of NPA on fundus fluorescein angiography (FFA) in DME. The study included 3,014 FFA images of 221 patients with DME. We use 3 convolutional neural networks (CNNs), including DenseNet, ResNet50, and VGG16, to identify non-perfusion regions (NP), microaneurysms, and leakages in FFA images. The NPA was segmented using attention U-net. For DR lesion classification, area under the curve is 0.8855 for NP regions, 0.9782 for microaneurysms, and 0.9765 for leakage classifier. The average precision of NP region overlap ratio is 0.643. NP regions of DME in FFA images are identified based on a new automated deep learning algorithm. This study serves as a theoretical basis for the application of intelligent guided laser.
Introduction
Diabetic macular edema (DME) remains a leading cause of vision impairment. Laser photocoagulation is a standard treatment, yet its success depends on the precise identification of pathologies such as microaneurysms, leakages, and non-perfusion areas (NPAs). NPAs are critical "dead zones" where blood flow has ceased. Currently, clinicians rely on Fundus Fluorescein Angiography (FFA) to manually delineate these regions—a process that is subjective, time-consuming, and prone to variability. There is a pressing clinical need for automated systems capable of spatial mapping to guide "intelligent" laser treatments.
Methodology
The researchers developed a two-stage deep learning pipeline to automate the detection and mapping of retinal lesions.
Multi-label Classification
The first stage involves identifying the presence of specific pathologies within FFA images. The study compared three Convolutional Neural Network (CNN) architectures: DenseNet, ResNet50, and VGG16. These models were trained to perform multi-label classification, identifying NP regions, microaneurysms, and leakages simultaneously.
Semantic Segmentation
To provide the precise spatial coordinates required for laser therapy, the team implemented an Attention U-Net for semantic segmentation of NPAs. Unlike standard U-Nets, the Attention U-Net utilizes "attention gates" to suppress irrelevant background noise and focus on the irregular features of non-perfusion regions.
Data and Preprocessing
The models were trained on a dataset of 3,014 FFA images from 221 patients. Ground truth was established through independent annotations by three ophthalmologists, requiring a consensus for labeling. To enhance feature visibility, images underwent Contrast Limited Adaptive Histogram Equalization (CLAHE) before processing.
Results and Discussion
DenseNet emerged as the superior architecture for classification, likely due to its dense connectivity which excels at capturing overlapping retinal pathologies.
Classification Performance (DenseNet): * NP Regions: AUC = 0.8855 * Microaneurysms: AUC = 0.9782 * Leakage: AUC = 0.9765
In the segmentation task, the Attention U-Net achieved an average precision of 0.643. While segmenting irregular NPAs is inherently challenging, the model's probability maps closely aligned with expert manual delineations, demonstrating its potential for clinical utility.
Conclusion
This study demonstrates that deep learning can effectively identify and segment critical retinal lesions in DME. By moving beyond simple screening to precise spatial mapping, this work provides a foundation for automated, intelligent guided laser treatments. Future efforts will focus on automating image registration between FFA scans and live surgical views, as well as exploring weakly supervised learning to reduce the burden of manual annotation.
Summary updated