糖尿病视网膜病变是糖尿病最常见也是最严重的并发症之一。为提高对糖尿病视网膜病变严重程度的诊断准确率，进一步为糖尿病视网膜病变治疗的精准用药提供依据，提出一种新型的特征融合网络模型。首先利用轻量化网络EfficientNet-B0提取眼底图像的不同层特征，使用高层特征结合三个不同空洞率的空洞卷积形成多尺度特征。然后引入多尺度通道注意力模块（MS-CAM），赋予高层特征和低层特征新的权重，对高低层特征进行融合，形成最终的特征表征，从而完成对糖尿病视网膜病变严重程度的分类。实验结果表明，所提模型的分类准确率达85.25%，表明其具有较好的可行性，在临床上给医生诊断起到了辅助作用，能更有效地预防糖尿病视网膜病变的进一步恶化。

糖尿病不仅会增加视网膜血管疾病的风险，严重时甚至会发展成为糖尿病视网膜病变。糖尿病视网膜病变的4种典型病理特征是微动脉瘤、出血、硬性渗出物和软性渗出物。随着机器学习尤其是深度学习的发展，智能辅助诊断医疗已经成为一种趋势，智能辅助诊断的前提是可以定性定量地提取出相应的病变区域。提出了一种基于深度学习级联架构参数优化的眼底病变检测模型，该模型有效解决了眼底病变的多尺度和小目标问题，在DDR数据集上检测病变的综合测试精度达0.380，检测性能优于目前主流的检测网络。

眼底照相是获取眼部图像的主要技术之一。利用眼底相机对视网膜病变区域进行拍摄可以获得清晰的图像，从获取的图像中能够直接观察到眼球中的渗出物、出血点和微血管瘤，根据检测出的病灶类型、数量和位置等信息可进行糖尿病视网膜病变分类。基于此，本文利用深度神经网络对糖尿病视网膜病变进行自动分类识别，提出了一种体系结构简单、在通用设备上运行速度快的卷积神经网络CA-RepVGG(CA代表Channel Attention，RepVGG为现有模块)。利用单路极简结构的RepVGG模块替代复杂的可使用性较差的模块作为分类模型的主体部位，并选用高效通道注意力机制ECA替代压缩注意力机制SE，以此来提升模型对病变分级的能力。此外，本文还将CA-RepVGG模型与传统的分类模型VGG-16、Inception-V3、ResNet-50和ResNext-50模型进行了比较。从比较结果可以看出，虽然CA-RepVGG模型的参数量最大，但由于其是单分支结构，且只有3×3卷积块，因此它的模型复杂度并不高，分类速度很快，比另外4个模型中分类速度最快的ResNet-50还高出15.3%。另外，利用两个混淆矩阵展示了所提模型的分类结果，其在两个数据集上的准确度都超过了92.4%，精确度不低于91.6%，灵敏度在93.8%以上。从实验结果可知，所提模型不仅可对糖尿病视网膜病变进行分类，而且相比其他现有模型具有一定的优越性。若将该模型应用在临床上，可以提高专业眼科医生在眼科疾病上的诊断效率。

The aim of this study is to detect whether the quantitative textural features of optical coherence tomography angiography (OCTA) images can be used to detect the eyes in the early stage of diabetic retinopathy (DR) from eyes with diabetes and no DR (NDR). Textural features including fractal dimension, contrast, correlation, entropy, energy, and homogeneity were calculated from the OCTA images. The Student's t-test was performed to identify the textural features that can be able to detect DR in the early stage. The area under the receiver operating characteristic (AUROC) curves, sensitivity, and specificity were calculated between the study groups. Our results indicated that the fractal dimension in ICP and SVP and the correlation in SVC showed the statistical significance between mild NPDR patients and NDR patients. The ROC analysis results showed that the AUROC of the fractal dimension in ICP was 0.736 with 0.773 sensitivity and 0.700 specificity. The cutoff point in ICP was 2.616. The OCTA-based fractal dimension was able to discriminate diabetic eyes with early retinopathy from healthy and NDR with higher sensitivity and specificity. The OCTA-based correlation showed the power to differentiate the mild NPDR eyes from the normal healthy and diabetic eyes. These results suggest that texture-based features of OCTA have the potential to assist in the assessment of therapeutic interventions to prevent early DR in diabetic subjects.

The size and shape of the foveal avascular zone (FAZ) have a strong positive correlation with several vision-threatening retinovascular diseases. The identification, segmentation and analysis of FAZ are of great significance to clinical diagnosis and treatment. We presented an adaptive watershed algorithm to automatically extract FAZ from retinal optical coherence tomography angiography (OCTA) images. For the traditional watershed algorithm, "over-segmentation" is the most common problem. FAZ is often incorrectly divided into multiple regions by redundant "dams". This paper analyzed the relationship between the "dams" length and the maximum inscribed circle radius of FAZ, and proposed an adaptive watershed algorithm to solve the problem of "over-segmentation". Here, 132 healthy retinal images and 50 diabetic retinopathy (DR) images were used to verify the accuracy and stability of the algorithm. Three ophthalmologists were invited to make quantitative and qualitative evaluations on the segmentation results of this algorithm. The quantitative evaluation results show that the correlation coefficients between the automatic and manual segmentation results are 0.945 (in healthy subjects) and 0.927 (in DR patients), respectively. For qualitative evaluation, the percentages of "perfect segmentation" (score of 3) and "good segmentation" (score of 2) are 99.4% (in healthy subjects) and 98.7% (in DR patients), respectively. This work promotes the application of watershed algorithm in FAZ segmentation, making it a useful tool for analyzing and diagnosing eye diseases.

Photobiomodulation (PBM) promoting wound healing has been demonstrated by many studies. Currently, 630 nm and 810 nm light-emitting diodes (LEDs), as light sources, are frequently used in the treatment of diabetic foot ulcers (DFUs) in clinics. However, the dose–effect relationship of LED-mediated PBM is not fully understood. Furthermore, among the 630 nm and 810 nm LEDs, which one gets a better effect on accelerating the wound healing of diabetic ulcers is not clear. The aim of this study is to evaluate and compare the effects of 630 nm and 810 nm LED-mediated PBM in wound healing both in vitro and in vivo. Our results showed that both 630 nm and 810 nm LED irradiation significantly promoted the proliferation of mouse fibroblast cells (L929) at different light irradiances (1, 5, and 10mW/cm2. The cell proliferation rate increased with the extension of irradiation time (100, 200, and 500 s), but it decreased when the irradiation time was over 500 s. Both 630 nm and 810 nm LED irradiation (5mW/cm2 significantly improved the migration capability of L929 cells. No difference between 630 nm and 810 nm LED-mediated PBM in promoting cell proliferation and migration was detected. In vivo results presented that both 630 nm and 810 nm LED irradiation promoted the wound healing and the expression of the vascular endothelial growth factor (VEGF) and transforming growth factor (TGF) in the wounded skin of type 2 diabetic mice. Overall, these results suggested that LED-mediated PBM promotes wound healing of diabetic mice through promoting fibroblast cell proliferation, migration, and the expression of growth factors in the wounded skin. LEDs (630 nm and 810 nm) have a similar outcome in promoting wound healing of type 2 diabetic mice.

针对糖尿病性视网膜图像数据集的不均衡、组织形态的特征提取不充分、分级准确率不高等问题,本文提出一种基于DR-Net模型的改进识别算法,即Improved DR-Net。选用Kaggle失明检测竞赛数据集APTOS 2019 Dataset,采用多种数据增强策略扩充数据集,并引入Eye-PACS数据集进行无偏修正,同时采用高斯滤波等形态学方法增强眼底图像特征;对ResNext50聚合残差结构进行预训练,通过迁移学习对基线模型进行参数及结构微调;引入空洞卷积代替普通卷积,融合注意力机制进一步优化模型性能。测试结果表明,本文所提的Improved DR-Net模型大大提高了糖尿病视网膜病变分级的准确率:阳性预测值97.9%,阴性预测值98.03%,准确率达到98.04%,远高于同类算法。结合深度学习技术辅助视网膜病变的筛查,对于视网膜病变的早期自动筛查具有一定的指导意义。

糖尿病性黄斑水肿（DME）是导致失明的主要原因之一, 由专业的医生通过检查光学相干扫描（OCT）图像是主要的诊断方法, 但这一过程不仅耗时而且容易误判, 提出一种辅助诊断模型来区分DME和正常黄斑。对原始OCT图像进行降噪、展平、裁剪预处理, 得到易于分类的病灶区图像, 在小波分解金字塔模型的基础上用局部二值模式方法对原图和低频子图像提取纹理特征; 与提取细节图像的灰度-梯度共生矩阵特征融合形成最终的全局特征, 并对其进行降维; 用weka平台的序列最小优化模型进行分类。在杜克大学数据集和临床数据集上的试验结果表明, 算法在两个数据集上验证的准确率分别为95.7%、95.3%, 灵敏性分别为95.3%、95.5%, 特异度分别为96.0%、95.1%。因此, 所提方法能有效对OCT图像分类, 为临床上视网膜疾病辅助诊断提供技术支撑。