頁籤選單縮合
題 名 | PET-MRI Image Registration and Fusion Using Artificial Neural Networks |
---|---|
作 者 | Wang,Woei-fuh; Ngo,Frank QH; Chen,Jyh-cheng; Huang,Ray-ming; Chou,Kuo-liang; Liu,Ren-shyan; | 書刊名 | 醫學工程 |
卷 期 | 15:3 2003.06[民92.06] |
頁 次 | 頁95-99 |
分類號 | 410.1644 |
關鍵詞 | PET; MRI; Image registration; Image fusion; Artificial neural network; |
語 文 | 英文(English) |
英文摘要 | Multi-modality image registration and fusion play an increasingly important role in medicine. 3-D image reconstruction, object recognition, and medical imaging analysis, are just a few examples. Medical images provide essential information for clinical diagnosis. Good image quality can yield more accurate patient information, which can then be used for better clinical decision making. X-ray computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET), and single photon emission computed tomography (SPECT) are clinically established imaging modalities. Among them, CT and MRI images are anatomical imaging with high spatial resolutions. However their physiological information is limited. On the other hand, although SPECT and PET can provide physiological information, both spatial resolutions are too poor to provide clear anatomical information. Thus, it should be advantageous to combine images from multi-modality images so that the resulting image can provide both physiological and anatomical information with high spatial resolution for use in clinical diagnosis and therapy. In order to generate an overlapping image from different-modality images, image registration and various fusion techniques have been employed. In the past studies, artificial neural network (ANN) has been used in registration and fusion of multi-modality remote sensing data for the reconstruction of 3-D models of terrain regions and medical images. In this study, we apply multilayer back propagation (BP) and conjugate gradient back propagation (CGBP) NN model to train PET and MRI images. We use Matlab and Borland C++ Builder to implement the network training frame and find out the optimal training model for accurate image registration. Further, we evaluate the accuracy of the image registration using NN by phantom images. Our study shows that, the registration errors are less than 5mm, and using appropriate opacity is better to show the fusion image than using interleaved pixels with independent color scale. Finally, we use Borland C++ Builder to implement a user-friendly window-based interface for image registration and fusion system. |
本系統中英文摘要資訊取自各篇刊載內容。