Diabetic macular edema (DME) is the major cause of vision loss in the patients with diabetic retinopathy. Increasing prevalence of DME is tightly correlated with the global epidemic of diabetes mellitus (1, 2). DME is usually caused by the rupture of retinal barrier and increased permeability of retinal vessels, which is characterized by the leakage of fluid and other plasma components. The effusion can accumulate in the macula, resulting in edema (3, 4). In the clinical work, the presence and severity of retinopathy are required to be determined according to the size of edema area.

Optical Coherence Tomography (OCT) is a non-contact, non-invasive, and highly sensitive ophthalmic imaging method, which can enable early detection of diabetic macular edema by observing the transverse section of macular degeneration (5). Normal OCT image is shown in Figure 1A and OCT image with DME is shown in Figure 1B. DMEs accumulated in typical relative positions within the main retinal layers. Based on OCT patterns of DME, DME can be classified into three different patterns, including diffuse retinal thickening (DRT), cystoid macular edema (CME), and serous retinal detachment (SRD). CME normally starts to manifest symptoms in the inner retina, while SRD and DRT typically appear in the outer retina. In the severe advanced stages of DR, CMEs can also proliferate from the inner to the outer retina and merge with DRT (6). Thus, rapid and accurate detection of all types of edemas is of great significance for evaluating the progression of diabetic retinopathy. In the clinical work, DME is usually segmented by the well-trained experts (7). However, manual extraction of DME edemas is time-consuming and labor-intensive. Moreover, there is inevitable variability in the extraction results by different experts. With increased prevalence of diabetes, an increasing number of patients require disease management based on OCT images in the clinical practices. Thus, it is highly required to design an automatic method for rapid and accurate detection of DME in OCT images.

Image extraction is processed and analyzed according to the features, including image color, spatial structure, and texture information (8). Image extraction models can divide an image into several specific regions, such as threshold-based extraction model (9, 10), region-based extraction model (11, 12), and edge detection-based extraction model (13). With the development of deep learning, several models have been developed to extract DME, such as fully convolutional network (FCN), U-Net, and PSPNet. Based on these deep learning models, several scholars have also developed the improved models for DME extraction. Table 1 showed the strengths and weaknesses of different models for DME extraction.

The sources of OCT images are diverse and the quality of OCT images is not always stable. Moreover, the size and distribution of DMEs are not uniform and the borders of DMEs are blurred. Thus, it is still required to design a novel model to improve the accuracy of DME extraction in OCT images. In this study, we proposed a lightweight automatic model (DME-DeepLabV3+) based on the DeepLabV3+ architecture. The major contributions of the proposed DME-DeepLabV3+ are shown below:

The flowchart of the proposed model, DME-DeepLabV3+, was shown in Figure 2.

Low-level features of DME extraction by MobileNetV2; High-level features of DME extraction by the improved ASPP; Fusion and refinement of low-level and high-level features of DME by the decoder.

The hardware configurations used for this study are shown below: Windows 10, NVIDIA GeForce RTX 3060. The software environment is the deep-learning framework PyTorch 1.10.0, CUDA 11.3, and the programming language Python 3.9.

Seven metrics were calculated to estimate the extraction performance of DME-DeepLabV3+, including pixel accuracy (PA), mean pixel accuracy (MPA), precision (Pre), recall (Re), F1-score (F1), mean intersection over union (MIoU), and frames per second (FPS).

TP, FP, and FN denote the true positive region, false positive region, and false negative region, respectively. p i i is the number of edema area pixels which was correctly classified as edema areas; p i j is the number of background area pixels which are misclassified as edema areas; p j i is the number of edema area pixels which are incorrectly classified as the background; k is the labeling results of different classes, where k = 0 expressed as background class and k = 1 as DME class; f r a m e N u m is the number of OCT images that are input to the model when performing inference; e l a p s e d T i m e is the time consumed by the model when performing inference. PA is the overall pixel accuracy. MPA is the average pixel accuracy of DME and background. Pre and Re are the proportion of real DME regions in the samples predicted as DME and the proportion of correct predictions in all DME, respectively. F1-score (F1) is a balanced metric and determined by precision and recall. MIoU is a metric to measure the similarity of ground truth and prediction. FPS is the number of OCT images inferred per second.

To evaluate the performance of DME extraction of DME-DeepLabV3+ model, two comparative experiments were performed. In experiment 1, DME extraction performance of DME-DeepLabV3+ model was evaluated by comparing against DeepLabV3+ model under different settings. In experiment 2, DME extraction performance of DME-DeepLabV3+ model was evaluated by comparing against other end-to-end models, including FCN, UNet, PSPNet, ICNet, and DANet.

With the increased incidence of diabetes, DME has become a major cause of visual impairment in diabetic patients (28). DME occurs as a result of the disruption of blood-retinal barrier and consequent increase in vascular permeability (29). OCT allows longitudinal, functional, and microstructural analysis of human macula (30). However, manual labeling DME is time-consuming and labor-intensive (31). Automatic extraction of DME based on machine learning can help physicians assess disease severity, determine treatment options, and improve life quality of patients (32). Thus, it is urgent to develop an efficient model for DME detection. In this study, we proposed a lightweight model based on DeepLabV3+, termed DME-DeepLabV3+, to extract DME in OCT images. MobileNetV2 architecture was used as the backbone to extract the low-level features of DME and reduce the model complexity to enhance DME detection accuracy. With the help of improved ASPP structure, DME-DeepLabV3+ avoided the grid effects and learned more local information. Finally, the decoder was used to fuse the low-level and high-level features of DME and refined the results of DME extraction.

OCT image modality has been widely used for detecting DME due to its non-invasive and high-resolution features. Considering the clinical characteristics that are present in OCT images such as thickness, reflectivity or intraretinal fluid accumulation, DMEs have been categorized into three different types: SRD, DRT, and CME. Traditional DME detection is based on the low-level hand-crafted features, which require significant domain knowledge and are sensitive to the variations of lesions. Given great variability of morphology, shape, and relative ME position, it is difficult to detect all three ME types simultaneously. Our proposed model can achieve automatic and simultaneous detection of all three types of ME (SRD, DRT, and CME) in the ophthalmological field. However, the accuracy of DRT detection is still not good as SRD or CME detection. DRT is characterized by a sponge-like retinal swelling of the macula with reduced intraretinal reflectivity. In addition, DRT is characterized by uniform thickening of inner retinal layers but without macroscopic optical empty spaces. Thus, further improvement of our proposed model is still required for enhancing the accuracy of the automatic detection of DRT edemas.

In clinical practice, layer segmentation and fluid area segmentation can provide qualitative information and visualization of retinal structure, which is important for DME assessment and monitoring. Although commercial OCT devices with on-board proprietary segmentation software are available, the definition of retinal boundaries varies between the manufacturers, making the quantitative retinal thickness difficult. In addition, proprietary software is difficult to be used for image analysis from other OCT devices, which poses a great challenge for effective diagnosis of DME (33). Although automated methods for layer segmentation have been proposed, most of them usually ignore the priority of mutually exclusive relationships between different layers, which can also affect the accuracy of DME assessment (34). In future study, we will improve our model to consider both layer segmentation and fluid area segmentation for better monitoring the progression of DME in retinal diseases.

Both microaneurysm (MA) formation and DME lesions are the important signs of DR. Early and accurate detection of DME and MAs can reduce the risk of DR. Due to the small size of MA lesions and low contrast between MA lesion and retinal background, automated MA detection is still challenging. Many imaging modalities have been used to detect MAs, including color fundus images, optical coherence tomography angiography (OCTA), and fluorescein fundus angiography (FFA). However, MAs are situated on the capillaries, which are not often visible in color fundus images. Although FFA can capture the small changes of retinal vessels, FFA is an invasive method compared with other imaging modalities. OCTA can provide the detailed visualization of vascular perfusion and allow for the examination of retinal vasculature in 3D (35). In future study, we would also improve our model by considering the segmentation of FFA for better monitoring the progression of DME in retinal diseases. We would design modules with better feature extraction capabilities, such as embedding attention mechanism to the model, strengthening key information, suppressing useless information, and better capturing contextual information, to improve the generalization of the model for the diagnosis of retinal diseases.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

BY, ZW, and QJ were responsible for the conceptualization and data collection. BY, ZW, and YB were responsible for the experiment design and manuscript writing. JL and LS conducted the data collection and data entry. BY and ZW were responsible for overall supervision and manuscript revision. All authors contributed to the article and approved the submitted version.

This research was generously supported by the grants from the National Natural Science Foundation of China (Grant Nos. 82171074 and 82070983).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer ZH declared a shared parent affiliation with the authors LS, QJ, and BY to the handling editor at the time of review.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.