Texture Analysis for Linear Spectral Unmixing of Brain MR Image Classification

• Main Authors: Hung-Che Li, 李宏哲
• Other Authors: Yen-Chieh Ouyang
• Format: Others
• Language: en_US
• Online Access: http://ndltd.ncl.edu.tw/handle/20968995637042861780

碩士 === 國立中興大學 === 電機工程學系所 === 97 === Linear spectral unmixing (LSU) has recently applied to MR image classification and shown potential in MR image classification. Unlike the traditional classification which is mainly focused on inter-pixel correlation among data samples the LSU explores intra-pixel correlation to characterize spectral properties for classification. As a result, a major strength of the LSU is to perform mixed pixel classification by estimating abundance fraction of each tissue substance present in a pixel to provide the likelihood of each tissue substance to be classified in one particular class. This task cannot be accomplished by classical spatial domain-based classification techniques which perform pure-pixel classification by confusion matrices that involve hard decisions instead of soft decisions made by mixed pixel classification. Since the LSU is an intra-pixel-based technique and does not take advantage of inter-pixel correlation as the traditional image processing techniques do. In order for the LSU to be able to do so, texture analysis is included as a post-processing technique that can be used in conjunction with LSU to perform both spectral/spatial (i.e., intra-pixel and inter-pixel) analysis. Specifically, the gray level co-occurrence matrix (GLCM) is used as a base to generate various texture features to be used as training samples for a follow-up back propagation-based neural network for classification. To demonstrate the performance of our texture-based LSU approach, experiments are conducted for performance analysis and evaluation.