Classifying digital X-ray images into different human body parts

In medical information retrieval research, automatically classifying X-ray images based on body-parts is a challenging problem. In ImageCLEF's 2015 campaign there was a contest where the participants were challenged to cluster X-ray images into different groups based on presence of particular body-part in that X-ray image. In brief the challenge was to classify given X-ray images primarily into five groups which were: head-neck, body, upper-limb, lower-limb and true-negative. In our approach to solve this task we extracted features from the given images using dense multi-scale SIFT, used Elkan k-means clustering to create visual dictionary of extracted features, Randomized KD-Tree to speed up the Elkan K-Means, Spatial Pyramid Histogram as image descriptors to train Chi-kernel based SVM as classifier. Proposed system is able to classify X-ray images into one of the five classes with 85 percent accuracy.