Multimodality Brain Image Fusion Methods For Glioma Classification

Medical image fusion refers to extracting and then combining the meaningful information from multimodal source images and aims to produce a single fused image that is more informative and beneficial for clinical purposes and computer-aided diagnosis(CAD).So far,significant fusion schemes have been proposed and implemented.However,these schemes integrate maximum information of the source images at the cost of high sensor noise,well reflect the salient features at the cost of local features,sufficiently improve the local and texture feature information at the cost of low intensity and color distortion,and capture maximum features details at the cost of high computational complexity.Following these limitations,this dissertation considered some improved proposed solutions.First,by using the full advantage of nonsubsampled shearlet transform(NSST),the source images are decomposed in low(LFS)and high-frequency subbands(HFS).Local feature-based fuzzy pixel fusion rules are applied to fuse the LFS.The novel summodified Laplacian(NSML)based max-selection fusion rules are adopted for the fusion of HFS.Combining both these fusion strategies significantly enhances the local features such as points,edges,or contours and avoids the blur effect of the resultant fused image.The second algorithm extends the first for pseudo-color imaging modality fusion.The pseudo-color source images are first converted to YIQ color space,and then the Y(luminance component)image is decomposed by NSST using direction window size [8,8,4,4] to avoid long execution.Here,the local spatial frequency(SF)and region energies(RE)are used as an activity measurement for the fuzzy interface system(FIS)for the fusion of LFS.The proposed scheme efficiently captures the local and texture features by intensively suppressing color distortion due to a better correlation between intensity and chromatic channels of YIQ.Third,the fast local Laplacian filter(FLLF)is first applied to enhance edge information and capture the actual geometry of the source images.Then the combination of parameter-adaptive pulse coupled neural network(PA-PCNN)and improved salience measures and matching factor(FW-SMF)fusion strategies are applied to fused HFS and LFS in the NSST and YUV domain,respectively.The algorithm not only enriches the details and contrast of the fused image but also suppresses the sensor noise artifacts and the edge halos.The fourth scheme demonstrates the effect of fusion on the brain tumor grade classification.First,region energy and entropy-based fuzzy pixel rules and WSML based PA-PCNN fusion strategies are developed in the multi-level edge-preserving filters(ML-EPF)domain to fuse MRI images.Then,the three most powerful and still efficient convolutional neural networks(CNN),i.e.Res Net152,Inception V3,and Dense Net201,are separately trained with the single-modality datasets(SMD)from Brats2020 and fused modality dataset(FMD)from the proposed fusion algorithm.The classification results obtained for FMD is increased by 5% on average compared to that of SMD in term of accuracy,sensitivity,specificity,precision and F1 score.Hence,it verifies that the fusion of multimodality medical images can help CAD and radiologists in glioma grade classification.Several experiments were conducted on the brain image datasets taken from the Harvard Medical School database for every proposed fusion scheme.The subjective analysis performed on the experimental results obtained verifies that the outcome of the suggested schemes is more apparent,having approved quality contrast resolution and stand-alone provide complete information about the actual edges and contours.Similarly,the results obtained for the objective analysis metrics are improved compared to the state-of-art methods.