Enhancing The Accuracy And Robustness Of Multimodal MRI Image Registration Using Hybrid Deep Learning Approaches And Multi-Scale Feature Fusion

Abstract

Purpose: The purpose of the study is to improve the accuracy and the robustness of the multimodal MRI image registration, where the significant variations in the intensity, the resolution, as well as the anatomical structures limit performance of the conventional and the deep learning-based methods. The research addresses the question of whether the hybrid deep learning framework, which integrates the local and the global feature modeling, can achieve the reliable alignment across the different MRI modalities while maintaining the robustness to the noise and the cross-modality variations.

Methods: The hybrid CNN–Transformer based image registration framework is proposed, which integrates the multi-scale feature extraction, the self-attention and the cross-attention mechanisms, the domain adaptation strategies, as well as the spatial transformer networks. The model captures the fine-grained local features as well as the long-range global dependencies. The experiments are conducted on the Osteoarthritis Initiative MRI dataset. The performance is evaluated using the overlap-based, the intensity-based, as well as the deformation-based metrics, which include the Dice Coefficient, the Jaccard Index, the Hausdorff Distance, the Mutual Information, the PSNR, as well as the Bending Energy. The ablation studies, the robustness tests under the noise, as well as the cross-modality evaluations are also performed.

Results: The proposed method achieves the superior registration performance compared to the baseline models as well as the state-of-the-art models, where the Dice Coefficient reaches 0.87 and the Jaccard Index reaches 0.78, while the Hausdorff Distance is reduced to 9.5 mm. The results remain stable under the noisy conditions and across the different MRI sequences, which confirms the strong generalization capability.

Conclusion: The hybrid CNN–Transformer framework provides the accurate, the robust, as well as the generalizable multimodal MRI registration. The integration of the multi-scale features, the attention mechanisms, as well as the domain adaptation significantly enhances the alignment precision.