Scholarlink

<p>Multimodal medical image fusion fuses different imaging modalities—MRI, CT, and PET, etc. into a single composite image to improve diagnosis accuracy and treatment planning. However, traditional fusion methods lack information preservation and noise reduction. In this study, a novel fusion model guided by Convolutional Neural Networks (CNNs) is proposed for addressing these issues, which incorporates Adaptive Channel Attention Mechanism, channel-wise Attention Regularization and attention loss to effectively deal with the limitations. We utilized two-branch 2D CNNs to extract high-level features from each imaging modality and propose an Adaptive Channel Attention Mechanism to dynamically focus the most informative feature maps updated dynamically across the channels. One of the strengths of our Method is that Attention Regularization allows the attention to diversify across the region of interests across the image, enforcing adequate fusion quality, and a loss term guides the network on which areas it should pay more attention in input images. The presented model is validated using several medical image datasets with higher performance scores in important metrics like Edge Information Retention (QAB/F) measuring edge preservation during processing, PSNR comparing signal strength to noise with higher values indicating better quality, Mutual Information (MI) gauges shared information between images, SSIM evaluates image similarity based on structure and contrast, Entropy indicates image complexity and Spatial Frequency describing image detail based on intensity changes, than conventional methods.</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

Enhanced multimodal medical image fusion using adaptive channel regularized attention mechanism with convolutional neural network

  • 期刊论文

  • 发表日期：

    2025年9月19日

查看全文
  • Satish Chaurasiya,
  • Neelu Nihalani

摘要

Multimodal medical image fusion fuses different imaging modalities—MRI, CT, and PET, etc. into a single composite image to improve diagnosis accuracy and treatment planning. However, traditional fusion methods lack information preservation and noise reduction. In this study, a novel fusion model guided by Convolutional Neural Networks (CNNs) is proposed for addressing these issues, which incorporates Adaptive Channel Attention Mechanism, channel-wise Attention Regularization and attention loss to effectively deal with the limitations. We utilized two-branch 2D CNNs to extract high-level features from each imaging modality and propose an Adaptive Channel Attention Mechanism to dynamically focus the most informative feature maps updated dynamically across the channels. One of the strengths of our Method is that Attention Regularization allows the attention to diversify across the region of interests across the image, enforcing adequate fusion quality, and a loss term guides the network on which areas it should pay more attention in input images. The presented model is validated using several medical image datasets with higher performance scores in important metrics like Edge Information Retention (QAB/F) measuring edge preservation during processing, PSNR comparing signal strength to noise with higher values indicating better quality, Mutual Information (MI) gauges shared information between images, SSIM evaluates image similarity based on structure and contrast, Entropy indicates image complexity and Spatial Frequency describing image detail based on intensity changes, than conventional methods.