A New Multi-Atlas Registration Framework for Multimodal Pathological Images Using Conventional Monomodal Normal...

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Radiology

Background:

• Multi-atlas registration (MAR) transfers information from atlases to new images for tasks like segmentation.
• Conventional MAR methods struggle with multimodal pathological images due to missing modalities and pathological influences.
• Existing methods fail to accurately register normal, monomodal atlases to complex pathological scans.

Purpose of the Study:

• To develop a novel MAR framework capable of handling multimodal pathological images.
• To overcome limitations of conventional MAR in routine clinical image-based diagnosis.
• To improve the accuracy and robustness of region of interest segmentation in pathological brain images.

Main Methods:

• Synthesizing multimodal normal atlases from monomodal normal atlases using deep learning.
• Employing a multimodal low-rank approach to recover normal-looking images from pathological scans.
• Performing multi-channel registration between synthesized multimodal atlases and recovered images.

Main Results:

• The proposed MAR framework demonstrated enhanced accuracy and robustness in brain region of interest segmentation.
• Experimental results showed significant improvements compared to state-of-the-art methods.
• The method effectively utilizes multimodal information and mitigates pathological influences.

Conclusions:

• The novel MAR framework successfully addresses challenges in registering atlases to multimodal pathological images.
• The approach leads to more accurate and robust medical image segmentation.
• This framework offers a promising solution for routine image-based diagnosis involving pathological data.