Scholarlink

<p>The COVID-19 pandemic has highlighted the need for reliable and fast diagnostic systems that go beyond the limitations of RT-PCR and single-image analysis. This study introduces a Texture-Encoded Early Fusion Model that combines Computed Tomography (CT) and Chest X-ray (CXR) images at the input stage. Unlike most existing multimodal approaches that rely on feature-level or late-fusion strategies, the proposed method performs channel-level early fusion by stacking handcrafted CT texture representations with CXR images before convolutional feature learning, enabling early cross-modal interaction while preserving modality-specific characteristics. Texture descriptors, including the Discrete Wavelet Transform (DWT), Gabor Filtering, Histogram of Oriented Gradients (HOG), Gray-Level Co-occurrence Matrix (GLCM), and Local Binary Patterns (LBP), are integrated with a pretrained VGG19 backbone to extract combined spatial and texture features. This early fusion strategy enhances sensitivity to disease-specific patterns such as ground-glass opacities and lung consolidations. A structured augmentation strategy is employed to address class imbalance and handle unpaired CT and CXR data. The proposed framework is evaluated on binary, 3-class, and 4-class COVID-19 classification tasks, achieving accuracies of 96.84%, 95.73%, and 94.13%, respectively, and consistently outperforming single-modality baselines.</p><p>Ablation studies confirm that each texture descriptor contributes to performance improvement, with wavelet-based fusion yielding the strongest results. Further analysis of fusion operators shows that learnable mechanisms, such as 1<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation>1 convolution, outperform fixed operators by more effectively exploiting cross-modal information. Overall, the proposed model achieves a favorable balance between accuracy, robustness, and efficiency, demonstrating that texture-based early fusion is a practical and effective approach for multimodal medical image classification and can be extended to other diagnostic 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

Early Fusion of Texture-Encoded CT with CXR Images for Pulmonary Disease Classification

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Gautami Shingan,
  • Amol Jagtap,
  • Sachin Chaudhary

摘要

The COVID-19 pandemic has highlighted the need for reliable and fast diagnostic systems that go beyond the limitations of RT-PCR and single-image analysis. This study introduces a Texture-Encoded Early Fusion Model that combines Computed Tomography (CT) and Chest X-ray (CXR) images at the input stage. Unlike most existing multimodal approaches that rely on feature-level or late-fusion strategies, the proposed method performs channel-level early fusion by stacking handcrafted CT texture representations with CXR images before convolutional feature learning, enabling early cross-modal interaction while preserving modality-specific characteristics. Texture descriptors, including the Discrete Wavelet Transform (DWT), Gabor Filtering, Histogram of Oriented Gradients (HOG), Gray-Level Co-occurrence Matrix (GLCM), and Local Binary Patterns (LBP), are integrated with a pretrained VGG19 backbone to extract combined spatial and texture features. This early fusion strategy enhances sensitivity to disease-specific patterns such as ground-glass opacities and lung consolidations. A structured augmentation strategy is employed to address class imbalance and handle unpaired CT and CXR data. The proposed framework is evaluated on binary, 3-class, and 4-class COVID-19 classification tasks, achieving accuracies of 96.84%, 95.73%, and 94.13%, respectively, and consistently outperforming single-modality baselines.

Ablation studies confirm that each texture descriptor contributes to performance improvement, with wavelet-based fusion yielding the strongest results. Further analysis of fusion operators shows that learnable mechanisms, such as 1 \(\times \) 1 convolution, outperform fixed operators by more effectively exploiting cross-modal information. Overall, the proposed model achieves a favorable balance between accuracy, robustness, and efficiency, demonstrating that texture-based early fusion is a practical and effective approach for multimodal medical image classification and can be extended to other diagnostic applications.