Scholarlink

<p>The proposed research aims to develop an efficient multimodal retinal image registration framework for the earlier diagnosis of Diabetic Retinopathy (DR). By integrating a lightweight neural network based on a modified MobileNet architecture with an enhanced RANSAC (Random Sample Consensus) algorithm, the framework improves registration accuracy, reduces computational cost, and enhances diagnostic performance by increasing sensitivity, specificity, and robustness against outliers. Accurate detection of DR is often hindered by the limitations of unimodal imaging techniques, which may not provide enough detailed information. Traditional methods for registering multimodal images, such as color fundus images and Optical Coherence Tomography (OCT) scans, face difficulties in aligning the images accurately due to the presence of outliers. These challenges lead to suboptimal image alignment, reducing the overall effectiveness of DR diagnosis. While existing methods suffer from poor outlier handling and computational inefficiency, the integration of the proposed lightweight MobileNet and enhanced RANSAC explicitly addresses these challenges by reducing false matches and achieving faster alignment. One of the main causes of eyesight loss in the globe is diabetic retinopathy. To stop its development, early identification is essential. By combining data from many imaging modalities, multimodal retinal imaging provides a more thorough retinal structures and increases diagnostic precision. However, because these multimodal images range in scale, orientation, and noise, accurately aligning them is a challenging process. The proposed framework utilizes a modified MobileNet-based lightweight neural network for feature extraction from retinal images. It is combined with an enhanced RANSAC algorithm that optimizes image alignment by rejecting outliers and fine-tuning transformation matrices. The registration algorithm was tested on a dataset of color fundus images and OCT scans to evaluate its performance. On the proposed registration achieved MSE = 0.0045, improving by ↓0.014 compared to MI-based registration, and ALD = 0.45 px (↓0.16 px vs. standard RANSAC). On the pooled fundus + OCT test set, DR classification produced an AUC of about 0.98, representing a + 0.03 improvement over non-registered baselines. The average runtime was 12.3 ms per picture pair on CPU and ~ 8 ms on GPU. The proposed method achieves 98.8% accuracy, 97.6% precision, 97.1% recall, 97.4% F1-score, and an AUC of around 0.98. These results demonstrate the effectiveness and reliability of the proposed method for comprehensive retinal image analysis. The results demonstrate that the proposed framework significantly enhances the accuracy of DR detection, making it a valuable tool for clinical applications.</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

Efficient Multimodal Retinal Image Registration for Diabetic Retinopathy Detection Using a Lightweight Neural Network and Enhanced RANSAC Algorithm

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月8日

查看全文
  • Young-Jin Jung,
  • I. Manimozhi,
  • Temesgen Engida Yimer

摘要

The proposed research aims to develop an efficient multimodal retinal image registration framework for the earlier diagnosis of Diabetic Retinopathy (DR). By integrating a lightweight neural network based on a modified MobileNet architecture with an enhanced RANSAC (Random Sample Consensus) algorithm, the framework improves registration accuracy, reduces computational cost, and enhances diagnostic performance by increasing sensitivity, specificity, and robustness against outliers. Accurate detection of DR is often hindered by the limitations of unimodal imaging techniques, which may not provide enough detailed information. Traditional methods for registering multimodal images, such as color fundus images and Optical Coherence Tomography (OCT) scans, face difficulties in aligning the images accurately due to the presence of outliers. These challenges lead to suboptimal image alignment, reducing the overall effectiveness of DR diagnosis. While existing methods suffer from poor outlier handling and computational inefficiency, the integration of the proposed lightweight MobileNet and enhanced RANSAC explicitly addresses these challenges by reducing false matches and achieving faster alignment. One of the main causes of eyesight loss in the globe is diabetic retinopathy. To stop its development, early identification is essential. By combining data from many imaging modalities, multimodal retinal imaging provides a more thorough retinal structures and increases diagnostic precision. However, because these multimodal images range in scale, orientation, and noise, accurately aligning them is a challenging process. The proposed framework utilizes a modified MobileNet-based lightweight neural network for feature extraction from retinal images. It is combined with an enhanced RANSAC algorithm that optimizes image alignment by rejecting outliers and fine-tuning transformation matrices. The registration algorithm was tested on a dataset of color fundus images and OCT scans to evaluate its performance. On the proposed registration achieved MSE = 0.0045, improving by ↓0.014 compared to MI-based registration, and ALD = 0.45 px (↓0.16 px vs. standard RANSAC). On the pooled fundus + OCT test set, DR classification produced an AUC of about 0.98, representing a + 0.03 improvement over non-registered baselines. The average runtime was 12.3 ms per picture pair on CPU and ~ 8 ms on GPU. The proposed method achieves 98.8% accuracy, 97.6% precision, 97.1% recall, 97.4% F1-score, and an AUC of around 0.98. These results demonstrate the effectiveness and reliability of the proposed method for comprehensive retinal image analysis. The results demonstrate that the proposed framework significantly enhances the accuracy of DR detection, making it a valuable tool for clinical applications.