Multimodal medical image analysis using deep learning registration and LWT-SVD fusion

Accurately identification of brain tumors is challenging due to the limitations of single-modality imaging, which often struggles to provide diagnostic information. Image registration addresses this challenge by ensuring precise alignment of multimodal images, facilitating effective fusion for early brain disease diagnosis. However, several traditional image registration methods often struggle with complex anatomy and spectral-spatial distortions, leading to poor fusion quality. To address this challenge, a dual-step strategy is proposed for multimodal medical images to assist doctors in diagnosing diseases. During the initial stage, source and template image alignment is performed by VGG-19 to extract the features. Subsequently, dynamic inlier selection to optimize the feature matching process, thereby improving the robustness of registration. Finally, thin plate spline interpolation is used to compute the affine parameters, achieving accurate registration of source image and template image. In the second stage, fusion scheme is developed by employing lifting wavelet transform (LWT) and singular value decomposition (SVD). Specifically, LWT is applied to decompose the input images into multi-level frequency bands, including low-frequency (approximation) and high-frequency (detail) components. The fused wavelet coefficients are then reconstructed using inverse LWT to generate a final fused image with improved visual clarity and reliability. The effectiveness of the dual-step strategy is extensively evaluated on both monomodal and multimodal medical images from standard and real-world datasets. The results highlight notable enhancements in performance metric RMSE reduced from 0.2886 to 0.10543, SSIM improved from 0.7238 to 0.9312, PSNR climbed from 58.9262 to 68.81256, and CC increased from 0.9284 to 0.99731.