BONE TEXTURE CHARACTERIZATION USING A NOVEL TEXTURE ANALYSIS TECHNIQUE FOR BETTER DETECTION OF OSTEOPOROTIC CASES

Abstract

Osteoporosis is an ailment of bones that become porous owing to lessened bone mass besides changes in the micro-architectural structure of the bone. Visual assessment of osteoporotic subjects through standard radiograph images is a non-trivial challenge for researchers doing research work in biomedical image processing and pattern recognition. The similitude, observed by a common physician while inspecting images of bone micro-architecture, between healthy and osteoporotic subjects makes classification even tougher. The current research is aimed at finding a better methodology, and proposing a better classification modality to distinguish between healthy and osteoporotic subjects. To understand an underlying machine learning model with potential to perform the said task, we first propose to use a set of statistical features followed by conventional machine learning methods to perform classification of the images. In order to extract the statistical features, the images are first subjected to Gabor filtering which performs a multiresolution analysis of the images. Later, we perform feature extraction using a minor variant of LBP followed by random projection for feature reduction and subsequent, classification. The results are encouraging, having room for improvement but the main outcome of this study is the potential provided by the Gabor filters’ Multi Resolution (MR) nature, and their ability to produce a set of scale and rotation invariant features. This outcome was carried forward to design a classification framework using Deep learning. In this context, we have designed a convolutional neural network having the ability to perform classification of bone radiographs. In this network, the first convolutional layer is used for performing Gabor filtering of the images; this layer is fixed and is not trainable. The main advantage that we have achieved by doing this is that the use of gradient images is known to perform much better rather than using the raw bone radiographs for the classification of bone radiographs. The extraction of better low-level features, is quite useful as it enhances the regions in the images which are contributing more to the classification of the images. Experiments have been conducted on a biomedical imaging challenge dataset (IEEE-ISBI 2014). Our results have validated that rather than increasing the depth of the network, iv extracting the gradient images from the raw images and using them in order to subsequently perform classification gives much better outcomes.