Scholarlink

<p>Diabetic Macular Edema (DME) is an eye disease characterized by the accumulation of fluid leaking from retinal blood vessels in the macula, the central part of the retina responsible for vision. The underlying cause of DME is a potentially blinding consequence of another diabetic eye disease called Diabetic Retinopathy (DR). Despite the increasing number of DM patients, manually detecting the presence of DME from retinal Optical Coherence Tomography (OCT) images used to screen macular lesions presents a significant challenge for specialists. However, since DME can lead to diabetic vision loss, early diagnosis and treatment are crucial. Therefore, we propose an algorithm to quickly detect DME pathology from retinal OCT images using Deep Learning (DL), which is becoming increasingly common in many medical fields, to assist clinicians. This study presents four different deep learning-based models for detecting DME status in the presence of retinal cystoid fluid in the macular regions. These models, with ResNet-18, ResNet-50, MobileNet, and Xception backbones, are individually popularly used in segmentation tasks based on the DeepLabv3 + Convolutional Neural Network (CNN) model. The proposed models were trained on a dataset consisting of 3090 retinal OCT images from DME subjects, available on Kaggle. Various preprocessing techniques were applied to the dataset before training to minimize noise from the OCT scans. Each pixel of the OCT images was labeled with a class representing a specific disease, and feature extraction was performed. 80% of the dataset was allocated for training to identify and diagnose DME, and the training data was amplified using various data augmentation techniques. The remaining 20% ​​was reserved as a test set to validate the performance of the models. Experimental results were evaluated on OCT images in the dataset using commonly used performance metrics in segmentation tasks such as accuracy, Intersection-over-Union (IoU), and Dice Coefficient (DC). The ResNet-18 model, which had the highest performance, achieved scores of 99.94%, 82.96%, and 90.69% on these metrics, respectively. Our proposed CNN algorithms achieved high accuracy for DME diagnosis. The proposed model for DM patients using OCT images can both prevent vision loss through accurate diagnosis and assist specialists in screening for DME.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Retinal Diabetic Macular Edema Detection in OCT Images Using DeepLabv3 + with ResNet-18 Architecture

  • 期刊论文

  • 发表日期：

    2026年2月27日

查看全文
  • Muharrem Balcı,
  • Ahmet Alkan

摘要

Diabetic Macular Edema (DME) is an eye disease characterized by the accumulation of fluid leaking from retinal blood vessels in the macula, the central part of the retina responsible for vision. The underlying cause of DME is a potentially blinding consequence of another diabetic eye disease called Diabetic Retinopathy (DR). Despite the increasing number of DM patients, manually detecting the presence of DME from retinal Optical Coherence Tomography (OCT) images used to screen macular lesions presents a significant challenge for specialists. However, since DME can lead to diabetic vision loss, early diagnosis and treatment are crucial. Therefore, we propose an algorithm to quickly detect DME pathology from retinal OCT images using Deep Learning (DL), which is becoming increasingly common in many medical fields, to assist clinicians. This study presents four different deep learning-based models for detecting DME status in the presence of retinal cystoid fluid in the macular regions. These models, with ResNet-18, ResNet-50, MobileNet, and Xception backbones, are individually popularly used in segmentation tasks based on the DeepLabv3 + Convolutional Neural Network (CNN) model. The proposed models were trained on a dataset consisting of 3090 retinal OCT images from DME subjects, available on Kaggle. Various preprocessing techniques were applied to the dataset before training to minimize noise from the OCT scans. Each pixel of the OCT images was labeled with a class representing a specific disease, and feature extraction was performed. 80% of the dataset was allocated for training to identify and diagnose DME, and the training data was amplified using various data augmentation techniques. The remaining 20% ​​was reserved as a test set to validate the performance of the models. Experimental results were evaluated on OCT images in the dataset using commonly used performance metrics in segmentation tasks such as accuracy, Intersection-over-Union (IoU), and Dice Coefficient (DC). The ResNet-18 model, which had the highest performance, achieved scores of 99.94%, 82.96%, and 90.69% on these metrics, respectively. Our proposed CNN algorithms achieved high accuracy for DME diagnosis. The proposed model for DM patients using OCT images can both prevent vision loss through accurate diagnosis and assist specialists in screening for DME.