Automatic classification of medical X‐ray images: hybrid generative‐discriminative approach. Issue 5 (1st July 2013)

• Record Type: Journal Article
• Authors: Reza Zare, Mohammad; Mueen, Ahmed; Awedh, Mohammad; Chaw Seng, Woo
• Is Part Of: IET image processing. Volume 7:Issue 5(2013)
• Page Start: 523
• Page End: 532
• Publication Date: 2013-07-01
• Subjects: image classification -- image representation -- medical image processing -- X‐ray imaging; image representations -- ImageCLEF 2007 medical database -- probabilistic latent semantic analysis -- discriminative classification -- classification performance -- hybrid generative‐discriminative approach -- medical X‐ray images -- automatic classification; Image processing -- Periodicals; 621.36705
• DOI: 10.1049/iet-ipr.2013.0049
• Languages: English
• ISSNs: 1751-9659

Abstract

Abstract : A new approach is presented to improve the classification performance of medical X‐ray images based on the combination of generative and discriminative classification approach. A set of labelled X‐ray images were given from 116 categories of different parts of body and the aim is to construct a classification model. This model was then used to classify any new X‐ray images into one of the predefined categories. The classification task started with extracting local invariant features from all images. A generative model such as probabilistic latent semantic analysis (PLSA) was applied on extracted features in order to provide more stable representation of the images. Subsequently, this representation was used as input to a discriminative support vector machine classifier to construct a classification model. The experimental results were based on ImageCLEF 2007 medical database. The classification performance was evaluated on the entire dataset as well as the class specific level. It was also compared with other classification techniques with various image representations on the same database. The comparison results showed that superior performance has been achieved especially for classes with less number of training images. Thus, only 7 out of 116 classes were left with accuracy rate below 60% as it differs from the results obtained using other classification approaches. This was attained by exploiting the ability of PLSA to generate a better image representation, discriminative for accurate classification and more robust when less training data are available. The total classification rate obtained on the entire dataset is 92.5%.