Medical Image Fusion Using Symmetric and Asymmetric Strategies: Evaluating Classical and Deep Learning Methods

By combining complementary data from various imaging modalities, including Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), multimodal medical image fusion significantly improves diagnostic accuracy. Despite notable progress, there are still no comprehensive assessment frameworks that fairly contrast cutting-edge deep learning methods with conventional symmetric and asymmetric fusion techniques. Using a wide range of models, such as Average Fusion, Principal Component Analysis (PCA), Discrete Wavelet Transform (DWT), Laplacian Pyramid, Saliency Mapping, Guided Filtering, and the DenseFuse Convolutional Neural Network (CNN) framework, this study suggests a unified comparative method for fusion of multimodal medical images. All techniques were applied to uniformly preprocessed CT-MRI datasets and assessed using a variety of quantitative metrics, including the Edge Preservation Index (EPI), Mean Squared Error (MSE), Entropy, Mutual Information (MI), Standard Deviation (STD), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM). The findings show that although deep learning-based methods such as DenseFuse provide strong feature extraction capabilities, they may have stability and generalizability issues in specific situations. However, asymmetric techniques provided better edge preservation and localized detail enhancement, while conventional symmetric techniques such as PCA and DWT demonstrated more consistent and interpretable performance across metrics. In addition to outlining the advantages and disadvantages of each category, this thorough analysis offers vital information to select the best fusion technique based on image properties, computational efficiency, and diagnostic requirements. The findings show that PCA and Guided Filtering produced superior SSIM and PSNR values (up to 0.9766 and 35.16 dB, respectively), while deep learning-based DenseFuse showed performance limitations under the given conditions, indicating domain adaptation issues. The results provide a strong basis for further investigation into multimodal image fusion and its useful implementation in clinical settings.