Scholarlink

<p>Integrating deep learning into medical imaging is a significant advancement in healthcare. There is continuous improvement in deep learning architectures, resulting in a better understanding of various medical conditions. Retinal image segmentation is an important area of research that provides clinical knowledge of the retina and diseases that manifest in it. We propose an improved and novel deep ResUNet architecture with Attention Aware Multi-Scale Feature Fusion Block (AMFF) for retinal vessel segmentation. We use the AMFF block to fuse the skip connections to collect global information and enhance this fused information using a channel attention mechanism to focus on informative channels. We use Alignment Gates equipped with dropout layers to hand out quality information to each level of the decoder. We also use a spatial attention module in decoding units to focus on the locations of the blood vessels. Residual connections in the encoder and decoder feature extraction blocks allow gradients to flow. We integrate the AMFF blocks in two ways and experiment on publicly available retinal fundus datasets, DRIVE, STARE, CHASE-DB1, and FIVES. Our results outperform those of other recent networks regarding accuracy, sensitivity, and AUC.</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

Deep ResUNet with attention-aware multi-scale feature fusion block for retinal vessel segmentation

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Ruhi R. Kabra,
  • Manesh B. Kokare

摘要

Integrating deep learning into medical imaging is a significant advancement in healthcare. There is continuous improvement in deep learning architectures, resulting in a better understanding of various medical conditions. Retinal image segmentation is an important area of research that provides clinical knowledge of the retina and diseases that manifest in it. We propose an improved and novel deep ResUNet architecture with Attention Aware Multi-Scale Feature Fusion Block (AMFF) for retinal vessel segmentation. We use the AMFF block to fuse the skip connections to collect global information and enhance this fused information using a channel attention mechanism to focus on informative channels. We use Alignment Gates equipped with dropout layers to hand out quality information to each level of the decoder. We also use a spatial attention module in decoding units to focus on the locations of the blood vessels. Residual connections in the encoder and decoder feature extraction blocks allow gradients to flow. We integrate the AMFF blocks in two ways and experiment on publicly available retinal fundus datasets, DRIVE, STARE, CHASE-DB1, and FIVES. Our results outperform those of other recent networks regarding accuracy, sensitivity, and AUC.