An Automatic Detection Model of Microaneurysms Based on Improved FC-Densenet

Diabetic retinopathy (DR) is a common eye disease, which leads cause of blindness all around the world. 40 Microaneurysms (MAs) is one of the early symptoms of DR. Accurate and effective MAs detection and 41 segmentation is an important step for the diagnosis and treatment of DR. In this paper, we propose an automatic 42 model for detection of MAs in fluorescein fundus angiography (FFA) images. The model mainly consists of two 43 steps. The first step is pre-processing of FFA images, where the quality of FFA images is improved by 44 Histogram Stretching and Gaussian Filtering algorithm. The second step is to detect MAs regions, where the 45 MAs regions are detected by improved FC-DenseNet. We compare the proposed model with traditional FC- 46 DenseNet model and other previously published models. The experimental result shows that our proposed model 47 has the highest scores on evaluation metrics of pixel accuracy ( PA ), mean pixel accuracy ( MPA ), precision ( Pre ), 48 recall ( Re ), F1-score ( F1 ) and mean intersection over union ( MIoU ), which are 99.97%, 94.19%, 88.40%, 49 89.70%, 88.98% and 90.14%, respectively. The result suggests that the performance of our proposed model is 50 closer to the ground truth of MAs detection. Our proposed model would be helpful for ophthalmologists to find 51 the symptoms more quickly and to take better treatment measures in the screening process of diabetic 52 retinopathy. 53


55
Retinal microaneurysms (MAs) is defined as small swelling of tiny blood vessels, which mainly locates in the 56 inner nuclear layer and deep capillary layer. It often occurs as early clinical signs of diverse retinal or systemic 57 diseases, including diabetic retinopathy (DR), retinal vein occlusions, infectious and so on. The number and 58 turnover of retinal MAs are considered as indicators to assess the presence, severity, and progression risk of 59 related retinopathy. Thus, early handling of MAs is needed to prevent vision loss caused by these retinopathies, 60 especially DR. MAs can be identified by many modern imaging technologies including color fundus photography, 61 fundus fluorescein angiography (FFA) and optical coherence tomography angiography (OCTA). Clinically, FFA 62 is well-recognized as the gold standard to visualize retinal vasculature and routinely used to describe the subtle 63 vascular alterations. 64 FFA is highly sensitive and demonstrates MAs as a hyperfluorescent dots in the early phase, contributing to 65 identification or evaluation of related retinal diseases. It is an important imaging modality, which captured after 66 intravenous injection of fluorescein dye using the dedicated fundus camera equipped with excitation and barrier filters. However, there are still some limitations in clinical applications. First, manual detection and quantification 68 of MAs are labor-intensive and time-consuming. With the increasing amount of FFA images that require analysis, 69 there is no sufficient number of ophthalmologists, especially in some rural areas. Second, such manual work is 70 subjective and error-prone, leading to poor reproducibility. Moreover, it is also infeasible for large-scale FFA 71 image analysis. 72 Therefore, the automated computer system may be helpful for ophthalmologists to identify and assess MAs 73 more efficiently. Numerous methodologies have been proposed currently. Zhang et al. [1] presented a novel MA 74 detection method based on a deep neural network with a multilayer attention mechanism for retinal fundus 75 images. Mazlan et al. [2] proposed an automatic detection of microaneurysms (MAs) in the fundus retina images. 76 Firstly, the images were filtered and the contrast enhanced. Then, the images were segmented using H-maxima 77 and thresholding technique. Long et al. [3] proposed a microaneurysms' detection method using machine learning 78 based on directional local contrast (DLC) for the early diagnosis of DR. Sarhan et al. [4] proposed a two-stage 79 deep learning approach for microaneurysms segmentation using multiple scales of the input with selective 80 sampling and embedding triplet loss. Yang et al. [5] proposed a method based on improved Hessian matrix 81 eigenvalue analysis to detect microaneurysms and hemorrhages in the fundus images of diabetic patients. Kou et Where 2  is the variance of Gaussian Filtering, l is the size of the filter kernel.

134
Here, the focal loss function is expressed as   Figure 3.and the results of evaluation metrics is shown in Table 1. 187 188 189  Table 1. Performance comparison between traditional FC-DenseNet models and our model.

Model
Evaluation metrics 193 Based on Figure 3 and Table 1

2 Compared with other end-to-end models 200
We compared our proposed model against other end-to-end models, including deeplabV3+ and PSPNet, to 201 evaluate the detection performance of MAs. The detection results of MAs are shown in Figure 4.and the 202 results of evaluation metrics is shown in Table 2

209
Based on Figure 4 and  Microaneurysms are the first clinically observable manifestations of diabetic retinopathy. Early diagnosis and 218 timely intervention can halt or reverse the progression of this disease [22]. In this paper, we propose a two-step 219 model, MAs-FC-DenseNet, for automatic detection of MAs in FFA images. Firstly, the pre-processing of FFA 220 images is conducted to enhance the contrast and to reduce the noise of FFA images. Then, MAs are detected by 221 our improved FC-DenseNet.
Most FFA images suffer from high noise and low contrast. Moreover, in FFA images, it is difficult to 223 distinguish MAs from blood vessels. Therefore, we improve the quality of FFA images by Histogram Stretching 224 and Gaussian Filtering in the pre-processing step. Due to the features of MAs, small and less in FFA images, it is 225 difficult to detect MAs accurately. Therefore, we use the FC-DenseNet model to detect the deep features of MAs, 226 and use the focal loss to enhance the detection accuracy of MAs in the detection step. 227 Compared with other deep learning network models, such as DeeplabV3+ and PSPNet, it could be found that 228 our proposed model achieves higher precision. The evaluation metrics, pixel accuracy ( The data used to support the findings of this study are included within the article. 244

Conflicts of Interest 245
The author(s) declare(s) that there is no conflict of interest regarding the publication of this paper. 246

Authors' Contributions 247
Biao Yan and Qing Jiang were responsible for the conceptualization, data collection. Zhen-Hua Wang and 248 Xiao-Kai Li were responsible for the experiment design and manuscript writing; Mu-Di Yao conducted the 249 data collection and data entry; Biao Yan and Zhen-Hua Wang were responsible for overall supervision and 250 manuscript revision 251