Scholarlink

<p>Since diabetic retinopathy (DR) is a major cause of avoidable blindness, accurate and easily implementable automated screening methods are required. This paper presents a unique lesion-aware DR detection mechanism using a Sparse Approximate Variational Autoencoder (SAVAE). This method is built to cultivate discriminative and sparse latent representations that emphasize clinically important retinal abnormalities. Reliable multi-class DR grading is further supported by the enhanced integration of structural and contextual features made feasible by the decoder reconstruction and feature-fusion technique. SAVAE enhances sensitivity to early-stage DR by prioritizing micro-lesions, such as microaneurysms and hemorrhages, by enforcing L1 sparsity inside the variational bottleneck. After training for 200 epochs, the model’s performance was evaluated using established assessment metrics. It was found that its performance was much better than that of traditional models like Vision Transformer, DMSN, VAE, and ML-Net. The results were remarkable: 95.4% accuracy, 94.5% precision, 99% recall and 97% F1-score. Furthermore, the ablation experiment showed that performance had a reduction of 2–3% when the sparsity constraint or the decoder reconstruction was removed, which proves their significance. The findings indicate that it can be applied in practical screening of DR, particularly in big community health institutions, where screening early is vital.</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

A Multi-stage Diabetic Retinopathy Classification System Based on Feature Fusion Using Sparse Approximate Variational Autoencoder

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Punam Sunil Raskar,
  • Vishal Naranje,
  • Swaroop Pillai,
  • Shital Ashok Pawar

摘要

Since diabetic retinopathy (DR) is a major cause of avoidable blindness, accurate and easily implementable automated screening methods are required. This paper presents a unique lesion-aware DR detection mechanism using a Sparse Approximate Variational Autoencoder (SAVAE). This method is built to cultivate discriminative and sparse latent representations that emphasize clinically important retinal abnormalities. Reliable multi-class DR grading is further supported by the enhanced integration of structural and contextual features made feasible by the decoder reconstruction and feature-fusion technique. SAVAE enhances sensitivity to early-stage DR by prioritizing micro-lesions, such as microaneurysms and hemorrhages, by enforcing L1 sparsity inside the variational bottleneck. After training for 200 epochs, the model’s performance was evaluated using established assessment metrics. It was found that its performance was much better than that of traditional models like Vision Transformer, DMSN, VAE, and ML-Net. The results were remarkable: 95.4% accuracy, 94.5% precision, 99% recall and 97% F1-score. Furthermore, the ablation experiment showed that performance had a reduction of 2–3% when the sparsity constraint or the decoder reconstruction was removed, which proves their significance. The findings indicate that it can be applied in practical screening of DR, particularly in big community health institutions, where screening early is vital.