PROJECT TITLE :

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

ABSTRACT:

For tasks like ROI segmentation, anatomical landmark recognition, and so on, information carried by atlases can be transferred to a new input image using multi-atlas registration (MAR). There are only normal anatomical structures in conventional MAR atlases. Consequently, the majority of MAR approaches are unable to process multimodal pathological pictures, which are commonly used in routine imaging-based diagnosis. There are two primary issues with registering monomodal atlases with normal appearances to multimodal pathological images: 1) missing imaging modalities in the monomodal atlases, and 2) the influence of pathological regions on the registration process. Here, we present an alternative approach for dealing with these issues. Deep learning-based image synthesisers are used to create multimodal normal atlases from monomodal normal atlases in this framework. A multimodal low-rank strategy to recover multimodal normal-looking images from multimodal pathological images is also proposed to lessen the impact of diseased regions. For the final step, the multimodal normal atlases can be mapped to the recovered multimodal pictures in a multi-channel approach. Our MAR framework is put to the test by segmenting multimodal tumour brain pictures based on their brain ROIs. Experimental results reveal that registration based on our method is more accurate and resilient than current state-of-the-art methods due to the use of multimodal information and the reduced influence from diseased regions.