A Robust Hybrid CNN–SURF Based Framework for Multimodal Brain Image Registration

Neurosurgeons and medical practitioners encounter challenges in accurately identifying brain tumors in preoperative and intraoperative images through visual examination alone. Although several conventional image registration methods have been developed to address this problem, their performance can be affected by noise, poor contrast, and mismatched feature points. To address these limitations, this work proposes a hybrid multi-stage registration framework that integrates a convolutional neural network (CNN) with Speeded-Up Robust Features (SURF) for medical image registration. Initially, the fixed and sensed images undergo preprocessing to improve image quality. Subsequently, multi-level feature representations are extracted using a pre-trained VGG19 network. Dynamic inlier selection with expectation maximization (EM) is then employed to iteratively refine feature correspondences and optimize the transformation. The optimized point correspondences are used as control points for thin-plate spline (TPS)-based nonrigid warping to generate an intermediate VGG19-registered image. Subsequently, SURF features are extracted from the fixed image and the intermediate registered image, and candidate feature correspondences are obtained using a brute-force (BF) matcher. Unreliable correspondences are further refined using the alpha-trimmed spatial relation correspondence (ATSRC) algorithm. Finally, a homography transformation is estimated from the refined correspondences to correct the remaining global/projective geometric misalignment and generate the final registered image. The proposed framework combines CNN-based hierarchical feature representation with SURF-based local feature detection and correspondence refinement to improve image alignment. The framework was evaluated on multimodal and monomodal medical image pairs from publicly available and clinical datasets, achieving RMSE of 10.1353 ± 3.1634, MSE of 111.7300 ± 65.7614, SSIM of 0.8850 ± 0.0275, PSNR of 35.5629 ± 1.8137 dB, and CC of 0.9614 ± 0.0379 across 10 publicly evaluated image-pair cases.