Scholarlink

Purpose <p>Diabetic retinopathy (DR) is a leading cause of vision loss worldwide. Traditional manual diagnosis by ophthalmologists is time-consuming and prone to delays. Deep learning (DL) models provide an automated approach to DR detection, enhancing early diagnosis and intervention. This study proposes an advanced method, SEMS DR Net, which integrates pre-trained ResNet models with Multi-scale Residual Blocks (MSRB) and the Squeeze and excitation (SE) attention mechanism, optimized through Bayesian optimization.</p> Methods <p>SEMS-DR Net is constructed using four ResNet variants ResNet-50, ResNet-101, ResNet-152, and ResNet-152V2 augmented with MSRB and SE modules. These models were trained and evaluated on three benchmark datasets: APTOS 2019, EyePACS, and DDR, targeting binary DR classification. Bayesian Optimization was employed to fine-tune model parameters for optimal performance.</p> Results <p>The ResNet152V2 + MSRB + SE model achieved superior performance across all datasets. On APTOS 2019, it achieved 98.74% accuracy, 98.91% precision, 98.76% recall, 98.80% F1-score, and a test loss of 0.034. On EyePACS, it recorded 96.65% accuracy, 96.80% precision, 96.60% recall, and 96.70% F1-score with a test loss of 0.045. On the DDR dataset, it attained 95.30% accuracy, 95.42% precision, 95.28% recall, and 95.32% F1-score, with a test loss of 0.061. The model also achieved a Cohen’s Kappa score of 0.9856 on APTOS, indicating excellent agreement with ground truth.</p> Conclusion <p>The proposed SEMS-DR Net demonstrates strong performance across the APTOS 2019, EyePACS, and DDR datasets, confirming its robustness and generalizability. Notably, it achieves the highest results on the APTOS dataset, indicating its suitability for training deep learning models due to its high-quality annotations. These findings provide valuable insights into the practical application of deep learning for DR detection, supporting early diagnosis and improved patient outcomes.</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

SEMS-DRNet: Attention enhanced multi-scale residual blocks with Bayesian optimization for diabetic retinopathy classification

  • 期刊论文

  • 发表日期：

    2025年10月6日

查看全文
  • Abini M.A,
  • S. Sridevi Sathya Priya

摘要

Purpose

Diabetic retinopathy (DR) is a leading cause of vision loss worldwide. Traditional manual diagnosis by ophthalmologists is time-consuming and prone to delays. Deep learning (DL) models provide an automated approach to DR detection, enhancing early diagnosis and intervention. This study proposes an advanced method, SEMS DR Net, which integrates pre-trained ResNet models with Multi-scale Residual Blocks (MSRB) and the Squeeze and excitation (SE) attention mechanism, optimized through Bayesian optimization.

Methods

SEMS-DR Net is constructed using four ResNet variants ResNet-50, ResNet-101, ResNet-152, and ResNet-152V2 augmented with MSRB and SE modules. These models were trained and evaluated on three benchmark datasets: APTOS 2019, EyePACS, and DDR, targeting binary DR classification. Bayesian Optimization was employed to fine-tune model parameters for optimal performance.

Results

The ResNet152V2 + MSRB + SE model achieved superior performance across all datasets. On APTOS 2019, it achieved 98.74% accuracy, 98.91% precision, 98.76% recall, 98.80% F1-score, and a test loss of 0.034. On EyePACS, it recorded 96.65% accuracy, 96.80% precision, 96.60% recall, and 96.70% F1-score with a test loss of 0.045. On the DDR dataset, it attained 95.30% accuracy, 95.42% precision, 95.28% recall, and 95.32% F1-score, with a test loss of 0.061. The model also achieved a Cohen’s Kappa score of 0.9856 on APTOS, indicating excellent agreement with ground truth.

Conclusion

The proposed SEMS-DR Net demonstrates strong performance across the APTOS 2019, EyePACS, and DDR datasets, confirming its robustness and generalizability. Notably, it achieves the highest results on the APTOS dataset, indicating its suitability for training deep learning models due to its high-quality annotations. These findings provide valuable insights into the practical application of deep learning for DR detection, supporting early diagnosis and improved patient outcomes.