Scholarlink

<p>Medical image segmentation is essential for clinical diagnosis, as it enables the identification of critical regions in CT scans, MRIs, and endoscopic images. However, conventional convolutional neural networks (CNNs) often fail to capture long-range dependencies and fine spatial details of images. This study introduces an improved hybrid architecture that integrates CNNs with Transformer-based self-attention mechanisms to enhance segmentation accuracy. In this study, we propose MLAT-Net, a hybrid architecture that integrates U-Net-style encoder-decoder networks with transformer-based self-attention and multi-level attention modules. We introduced two complementary components at the decoder stage: an Adaptive Scale Interaction Block (ASIB) for multi-scale feature aggregation and a Transformer-Aware Attention Core (TAAC) for modelling long-range dependencies, along with Position Attention Module (PAM) embedding at the bottleneck. The model is evaluated using ISIC-2018, Kvasir-SEG, and CVC-ClinicDB datasets. Comprehensive experiments with multiple optimizers and train/test splits (80:20, 70:30, and 60:40) demonstrated robust performance and generalization. MLAT-Net achieved superior Dice and IoU metrics compared to state-of-the-art models, whereas visualizations of decoder feature maps provided insights into model interpretability. Multi-level attention mechanisms effectively enhance feature representation. ensuring robust and generalizable performance across different medical imaging modalities. This study demonstrates that integrating CNNs with Transformer-based attention mechanisms significantly improves medical image segmentation and offer a reliable method for clinical applications.</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

MLAT -Net: A novel hybrid version of multilevel attention with transformer for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • Sachin B. Jadhav,
  • Prati Pal

摘要

Medical image segmentation is essential for clinical diagnosis, as it enables the identification of critical regions in CT scans, MRIs, and endoscopic images. However, conventional convolutional neural networks (CNNs) often fail to capture long-range dependencies and fine spatial details of images. This study introduces an improved hybrid architecture that integrates CNNs with Transformer-based self-attention mechanisms to enhance segmentation accuracy. In this study, we propose MLAT-Net, a hybrid architecture that integrates U-Net-style encoder-decoder networks with transformer-based self-attention and multi-level attention modules. We introduced two complementary components at the decoder stage: an Adaptive Scale Interaction Block (ASIB) for multi-scale feature aggregation and a Transformer-Aware Attention Core (TAAC) for modelling long-range dependencies, along with Position Attention Module (PAM) embedding at the bottleneck. The model is evaluated using ISIC-2018, Kvasir-SEG, and CVC-ClinicDB datasets. Comprehensive experiments with multiple optimizers and train/test splits (80:20, 70:30, and 60:40) demonstrated robust performance and generalization. MLAT-Net achieved superior Dice and IoU metrics compared to state-of-the-art models, whereas visualizations of decoder feature maps provided insights into model interpretability. Multi-level attention mechanisms effectively enhance feature representation. ensuring robust and generalizable performance across different medical imaging modalities. This study demonstrates that integrating CNNs with Transformer-based attention mechanisms significantly improves medical image segmentation and offer a reliable method for clinical applications.