Scholarlink

<p>Accurate and interpretable analysis of medical images requires the integration of complementary modalities such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT). This paper proposes <i>SE-MMFusionGAN</i>, a <i>Sustainable and Explainable Multi-Modal Fusion Generative Adversarial Network</i> that unifies interpretability, efficiency, and fusion quality within a single framework. Dual modality-specific encoders extract rich structural and textural representations, which are fused through a Cross-Modality Attention (CMA) mechanism to preserve clinically relevant features. An Activation-Efficiency Regularization module minimizes redundant activations, reducing computational and energy overhead, while a gradient-based attribution mechanism provides modality-aware explainability through interpretable heatmaps. The resulting fused representation enhances both visual interpretability for clinicians and downstream analysis, including anomaly localization. Extensive evaluations on benchmark datasets, including BraTS 2024, CT–MRI paired data, and BMAD, demonstrate that SE-MMFusionGAN achieves improved PSNR, SSIM, Dice, and IoU scores compared to state-of-the-art fusion networks, along with significant reduction in energy consumption. The framework thus establishes a balance between fidelity, interpretability, and sustainability, offering a practical step toward energy-aware and explainable multi-modal medical image fusion.</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

SE-MMFusionGAN: A Sustainable and Explainable Multi-modal Fusion GAN for MRI–CT Image Fusion and Anomaly Localization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月27日

查看全文
  • Pulkit Dwivedi,
  • Soumendu Chakraborty,
  • Niyaz Ahmad Wani,
  • Mansi Kajal,
  • Kuntal Gaur

摘要

Accurate and interpretable analysis of medical images requires the integration of complementary modalities such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT). This paper proposes SE-MMFusionGAN, a Sustainable and Explainable Multi-Modal Fusion Generative Adversarial Network that unifies interpretability, efficiency, and fusion quality within a single framework. Dual modality-specific encoders extract rich structural and textural representations, which are fused through a Cross-Modality Attention (CMA) mechanism to preserve clinically relevant features. An Activation-Efficiency Regularization module minimizes redundant activations, reducing computational and energy overhead, while a gradient-based attribution mechanism provides modality-aware explainability through interpretable heatmaps. The resulting fused representation enhances both visual interpretability for clinicians and downstream analysis, including anomaly localization. Extensive evaluations on benchmark datasets, including BraTS 2024, CT–MRI paired data, and BMAD, demonstrate that SE-MMFusionGAN achieves improved PSNR, SSIM, Dice, and IoU scores compared to state-of-the-art fusion networks, along with significant reduction in energy consumption. The framework thus establishes a balance between fidelity, interpretability, and sustainability, offering a practical step toward energy-aware and explainable multi-modal medical image fusion.