Parameter Optimization for Medical Image Modality Classification

Abstract

With increasing usage of medical images for the diagnosis in healthcare sector, the size of the image repository grows enormously. Image retrieval becomes a critical task with increasing size of repository. To address this problem, this article deals with the design of an automated system to predict the modality of medical image. This work then can be incorporated into image retrieval system with a large collection of medical images. Six modalities such as CT (computed tomography), XR (X-ray), PET (positron emission tomography), US (ultrasound), MR (magnetic resonance imaging) and PX (photograph) are considered in this experiment. Dense SIFT (scale-invariant feature transform) features, sampled at regular intervals, are extracted from the images, represented with bag-of-words histogram and classified by SVM (support vector machine). This paper explores three directions to improve the classification accuracy—usage of increasing number of training images, preferring spatial histogram rather than simple histogram and extending kernel map from linear to hellinger in SVM classifier. The obtained results are compared with existing complicated approaches and proved that better classification results are obtained with proposed simple approaches.