Document Type : Other

Author

• Khadijeh Aghajani

university of mazandaran

Abstract

Accurate multimodal image registration is a critical prerequisite in medical imaging, enabling reliable diagnosis and treatment planning. However, it remains challenging due to inherent intensity differences between modalities and spatially varying distortions. While existing methods address either modality conversion or illumination robustness separately, their combined treatment remains limited. In this research, we propose a two-stage framework that integrates deep learning-based modality transformation with distortion-robust mono-modal registration. First, a convolutional neural network converts the source modality to the target modality while preserving structural features through distortion-aware training. Second, a weighted Total Variation (wTV) registration method aligns the converted images, explicitly handling residual illumination variations. The proposed framework demonstrates consistent improvements across all test scenarios. It demonstrated substantially improved registration performance, supporting its applicability to multimodal clinical MRI alignment.

Keywords

• Nonrigid Multi-modal Image Registration
• Spatially Varying Intensity Distortion
• Deep Learning
• Modality Conversion

Main Subjects

• H.5. Image Processing and Computer Vision