Scholarlink

<p>Automatic and accurate medical image segmentation is crucial for the early diagnosis and treatment of cancerous diseases. Despite advancements in deep learning, precise lesion segmentation remains challenging due to significant variations in lesion size, position, and shape. While many deep learning models incorporate techniques such as data augmentation, multi-scale processing, and attention mechanisms, they often struggle with effectively capturing both global and local contextual information, leading to suboptimal segmentation performance in cases with extreme morphological variations. This paper proposes a deep learning model based on an encoder-decoder architecture, where a dual encoder integrates a transformer branch with ResNet to capture both local and global contextual information effectively. To enhance segmentation robustness across varying lesion sizes and shapes, the model incorporates multiple attention mechanisms. A dual spatial attention gate is introduced in skip connections to refine spatial feature selection, while an enhanced efficient channel attention module in the decoder recalibrates feature responses in a channel-wise manner. Additionally, a multi-scale attention module emphasizes critical feature maps across different scales, improving adaptability to lesion size variations. The proposed model is evaluated using five performance metrics to assess its effectiveness. The paper conducted comprehensive experiments on three datasets, namely ISIC2018 (skin), Endoscene (polyp), and BUSI (breast) to evaluate the efficacy of the proposed model. The proposed model achieved an IoU and Dice of 89.28% and 94.03% on ISIC2018, 91.90% and 91.18% on Endoscene, and 81.99% and 80.55% on the BUSI Dataset, respectively. The proposed model’s comparison with state-of-the-art models demonstrated its superiority in terms of performance metrics, generalizability test on ETIS-LaribPolypDB, Dataset B and visual results. Different components of the proposed model were validated through ablation studies. This paper proposes a deep learning model for variable-size lesion segmentation in 2D medical images. The architectural design, particularly the dual encoder and multi-scale attention modules, can be adapted for 3D medical images by replacing 2D convolutions with 3D convolutions and modifying attention mechanisms to process volumetric features. Future work will explore this extension with dedicated 3D medical imaging datasets.</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 dual encoder-decoder model with multi-attention for segmentation of variable size lesions in 2D medical images

  • 期刊论文

  • 发表日期：

    2026年6月24日

查看全文
  • Aasia Rehman,
  • Sajid Yousuf Bhat

摘要

Automatic and accurate medical image segmentation is crucial for the early diagnosis and treatment of cancerous diseases. Despite advancements in deep learning, precise lesion segmentation remains challenging due to significant variations in lesion size, position, and shape. While many deep learning models incorporate techniques such as data augmentation, multi-scale processing, and attention mechanisms, they often struggle with effectively capturing both global and local contextual information, leading to suboptimal segmentation performance in cases with extreme morphological variations. This paper proposes a deep learning model based on an encoder-decoder architecture, where a dual encoder integrates a transformer branch with ResNet to capture both local and global contextual information effectively. To enhance segmentation robustness across varying lesion sizes and shapes, the model incorporates multiple attention mechanisms. A dual spatial attention gate is introduced in skip connections to refine spatial feature selection, while an enhanced efficient channel attention module in the decoder recalibrates feature responses in a channel-wise manner. Additionally, a multi-scale attention module emphasizes critical feature maps across different scales, improving adaptability to lesion size variations. The proposed model is evaluated using five performance metrics to assess its effectiveness. The paper conducted comprehensive experiments on three datasets, namely ISIC2018 (skin), Endoscene (polyp), and BUSI (breast) to evaluate the efficacy of the proposed model. The proposed model achieved an IoU and Dice of 89.28% and 94.03% on ISIC2018, 91.90% and 91.18% on Endoscene, and 81.99% and 80.55% on the BUSI Dataset, respectively. The proposed model’s comparison with state-of-the-art models demonstrated its superiority in terms of performance metrics, generalizability test on ETIS-LaribPolypDB, Dataset B and visual results. Different components of the proposed model were validated through ablation studies. This paper proposes a deep learning model for variable-size lesion segmentation in 2D medical images. The architectural design, particularly the dual encoder and multi-scale attention modules, can be adapted for 3D medical images by replacing 2D convolutions with 3D convolutions and modifying attention mechanisms to process volumetric features. Future work will explore this extension with dedicated 3D medical imaging datasets.