Registration and Fusion of Multi-modal Medical Images

Obtaining medical images from different modalities is highly desirable for clinical diagnosis, radio therapy planning and detecting any abnormalities in the human body. To extract a great deal of information from multi-modal images and create a composite one, the images need to be registered and fused. This study explores accurate and highly robust registration and fusion approaches in the area of medical image analysis. In it, intensity-based similarity measures are used for both rigid and non-rigid registrations, and the fusion is accomplished by a pixel-level fusion scheme, with the performances of the proposed approaches evaluated on both simulated and clinical datasets.Firstly, the recently proposed sum of the conditional variance (SCV) similarity measure is applied on multi-modal medical image datasets with Gauss-Newton (GN) gradient-descent optimization using a 3D rigid body transformation and the registration results are compared with those from existing schemes. Although the SCV algorithms provide high robustness and a good convergence rate, their overall accuracy for the clinical dataset is not optimal. Therefore, a new approach, combining SCV with a dual-tree complex wavelet transform (DTCWT) mask, is used as a similarity measure with GN optimization. This new approach provides registered clinical multi-modal datasets with significantly improved registration accuracy.Secondly, a novel extension of the SCV similarity measure using partial volume interpolation (PVI) rather than quantized information is employed to make the joint histogram. The SCV(PVI) and SCV(PVI) with gradient mask algorithms are implemented for 3D non-rigid registration using GN optimization by applying affine and elastic transformations, respectively. The experimental results demonstrate that the SCV(PVI) and SCV(PVI) with gradient mask provide better registration results than existing algorithms.Finally, a multi-modal medical image fusion approach, DTCWT with principal component analysis (PCA), is proposed. In the DTCWT(PCA) algorithm, the input images are decomposed using DTCWT, their high- and low-frequency coefficients combined using the maximum fusion rule and PCA respectively, and then the inverse DTCWT applied to obtain the fused images. The qualitative and quantitative performance of this approach is compared with the experimental results from current fusion schemes and it is found that the DTCWT(PCA) method produces the best fusion results