Scholarlink

<p>Deep learning architectures have recently made outstanding progress, expanding their efficacy to various computer vision applications. The segmentation of retinal blood vessels from retinal fundus images has received a lot of attention among these applications since it is crucial for the early identification of many retinal illnesses, such as hypertension and diabetic retinopathy. The intricate structure of blood vessels coupled with variations in thickness, presents a formidable challenge in segmenting retinal images. While U-net-based approaches excel in medical image segmentation, their dependency on continuous pooling layers and convolution operations can compromise spatial and texture information. Although Transformers demonstrate comparable or superior performance to convolutional neural networks (CNNs), they demand extensive training datasets. While pre-training could address this issue, it introduces a significant computational overhead. Therefore, there is a need to devise a lightweight model to enhance efficiency in tackling these challenges. This paper introduces a Group Bottleneck Transformer U-Net (GBAT-UNet) enhanced with an Integrated Attention Module (IAM) and a Feature Adaptation Unit (FAU) designed to reduce the semantic gap between feature maps across skip connections. The proposed network demonstrates competitive performance in retinal vessel segmentation on three public benchmark datasets DRIVE, CHASE_DB1, and STARE. GBAT-UNet achieves an F1-score of 84.76% on DRIVE, 82.23% on CHASE_DB1, and 79.27% on STARE, along with high specificity values of 98.88% and 98.91% on CHASE_DB1 and STARE, respectively. These results highlight the model’s ability to accurately delineate retinal vasculature while effectively minimizing false-positive detections.</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

GBAT-UNet: group bottleneck transformer with attention in UNet for retinal vessel segmentation

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Ananya Bose,
  • Prerana Mukherjee,
  • Anasua Sarkar

摘要

Deep learning architectures have recently made outstanding progress, expanding their efficacy to various computer vision applications. The segmentation of retinal blood vessels from retinal fundus images has received a lot of attention among these applications since it is crucial for the early identification of many retinal illnesses, such as hypertension and diabetic retinopathy. The intricate structure of blood vessels coupled with variations in thickness, presents a formidable challenge in segmenting retinal images. While U-net-based approaches excel in medical image segmentation, their dependency on continuous pooling layers and convolution operations can compromise spatial and texture information. Although Transformers demonstrate comparable or superior performance to convolutional neural networks (CNNs), they demand extensive training datasets. While pre-training could address this issue, it introduces a significant computational overhead. Therefore, there is a need to devise a lightweight model to enhance efficiency in tackling these challenges. This paper introduces a Group Bottleneck Transformer U-Net (GBAT-UNet) enhanced with an Integrated Attention Module (IAM) and a Feature Adaptation Unit (FAU) designed to reduce the semantic gap between feature maps across skip connections. The proposed network demonstrates competitive performance in retinal vessel segmentation on three public benchmark datasets DRIVE, CHASE_DB1, and STARE. GBAT-UNet achieves an F1-score of 84.76% on DRIVE, 82.23% on CHASE_DB1, and 79.27% on STARE, along with high specificity values of 98.88% and 98.91% on CHASE_DB1 and STARE, respectively. These results highlight the model’s ability to accurately delineate retinal vasculature while effectively minimizing false-positive detections.