Abstract
2309
Objectives Image visual contents, such as textures and shapes, have been used as low-level visual features in retrieval of medical images. Such content-based image retrieval has potential to assist diagnosis of neurodegenerative disorders. However, it is challenging for physicians to interpret such low-level visual features, which are semantically meaningless. We proposed a new approach to translate image visual contents into semantic disorder features and evaluated its retrieval performance for multiple types of neurodegenerative disorders.
Methods Parametric images of rCMRGlc were spatially normalized to a standard atlas using SPM2. 3D discrete curvelets (DCvT) in 5 scales were used to extract low-level visual features for each rCRMGlc image. Fuzzy-C-means clustering algorithm was then applied to classify feature vectors into 12 semantic words, and each cluster/semantic word represents one type of disorder or normal group. The quadratic form distance was used in measuring the similarity among the semantic word histograms of subjects with the bag-of-words model. The method was compared with the traditional 3D gray level occurrence matrices (GLCM) algorithm using 209 FDG-PET neurological studies, which were acquired on a CTI ECAT 951R scanner and classified into 11 types of disorders and one normal group according to the diagnostic reports. The mean average retrieval precision was used in the evaluation.
Results Overall, our proposed semantic-word-based approach outperformed the traditional 3D GLCM algorithm by 2.38% with semantic word clustering and 2.04% without clustering, respectively; meanwhile the feature space was significantly reduced from 546 dimensions for the traditional 3D GLCM to 12 dimensions for the proposed approach.
Conclusions Our method substantially reduces feature dimensions for image retrieval, and is efficient in retrieving relevant neurodegenerative studies with meaningful semantic words related to disorder types. This may facilitate the retrieval of neuroimaging data for a wide range of neurological disorders.
Research Support ARC grants