Study On Multi-modality Biomedical Image Fusion Techniques Based On Multi-scale Geometric Analysis And Sparse Representation Theory

Nowadays,biomedical imaging techniques can show organisms’ structural and functional information in different scales such as organ,tissue,cell,molecule and gene.However,different imaging modalities have different applicable fields,advantages and dis advantages.Fusing biomedical images obtained from different imaging modalities and scales is a vitially important research interest in the whole field of biomedical imaging.Study on biomedical image fusion techniques has academic and clinical significances in diagnosing by clinical imageology,precisely locating lesion,designing plan of radiotherapy treatment,developing program of surgical operation and evaluating therapeutic effects.This dissertation conducted thorough research on multi-modality biomedical image fusion tasks.We presented four sub-tasks as follows.(1)For CT and MR image fusion,the fusion method combining nonsubsampled shearlet transform(NSST)with sparse representation(SR)was proposed.First,the CT and MR images were both transformed to nonsubsampled shearlet transform(NSST)domain.Then the high frequency subbands were merged by the absolute value maximum(AVM)rule while the low frequency subands were merged by a SR based approach.And the dynamic group sparsity recovery(DGSR)algorithm was first explored to implement the sparse coding of low frequency subands.Finally the fused image was obtained by performing the inverse NSST on the merged high frequency and low frequency subbands.The comparative experimental results demonstrated that the proposed fusion method could provide better performance in terms of subjective assessment,and acquire the best results for six metrics,the second best result for one metric and fine results for two metrics among nine objective evaluation metrics.(2)For PET/SPECT and MR image fusion,the fusion method based on NSST was proposed.First,the PET/SPECT and MR images were decomposed by NSST.Then the high frequency subbands were merged by AVM rule while the low frequency subands were merged by the proposed Haar wavelet-based energy(HWE)rule.Finally the fused image was obtained by performing the inverse NS ST on the merged high frequency and low frequency subbands.The comparative experimental results revealed that the proposed fusion method could provide satisfactory visual effects,and acquire the best results for four metrics and the second best result for one metric among five objective evaluation metrics.(3)For GFP and phase contrast image fusion,the fusion method based on convolutional sparse representation(CSR)was proposed.First,the GFP and phase contrast images were decomposed by CSR.Then the detail layers were merged by the sum modified Laplacian(SML)rule while the base layers were merged by the proposed adaptive region energy(ARE)rule.Finally the fused image was achieved by carrying out the inverse CSR on the merged detail layers and base layers.The experimental results reflected that the proposed fusion method could provide satisfactory visual effects,and achieve the best results for four metrics,the second best result for one metric and middling result for one metric among six objective evaluation metrics.Moreover,the proposed fusion method showed fine robustness to the mis-registration of GFP and phase contrast images.(4)For PET and CT image fusion,the fusion method combining CSR and NS ST was proposed.First,CSR was operated on PET and CT images to obtain base layers and detail layers.Second the base layers were decomposed by NSST to get low frequency and high frequency subbands.Then the detail layers and high frequency subbands were merged by the proposed local cross correlation maximum(LCCM)rule while the low frequency subbands were merged by the region energy maximum(REM)rule.Finally the fused image was acquired by performing the inverse NSST on the merged subbands and inverse CSR on the merged layers.The experimental results indicated that the proposed fusion method could provide good performance in terms of subjective assessment,and achieve the best results for five metrics and the second best result for one metric among six objective evaluation metrics.Moreover,the proposed fusion method exhibited fine robustness to the mis-registration of PET and CT images.This dissertation focused on the research of four kinds of multi-modality biomedical image fusion tasks.Four kinds of image fusion methods and three kinds of specially designed fusion rules were proposed according to the characteristics of different imaging modalities.The experimental results demonstrated that the proposed methods could provide good fusion performance in terms of subjective analysis and objective evaluation metrics.