Content Based Radiographic Images Indexing and Retrieval Using Pattern Orientation Histogram

Document Type : Original Paper


1 Ph.D. Student of Computer Engineering, Islamic Azad University, Research and Science Branch, Tehran, Iran

2 Assistant Professor and Director of Multimedia Research Group, IT Faculty, Iran Telecom Research Center, Tehran, Iran

3 Associate Professor and Director of IT Faculty, Iran Telecom Research Center, Tehran, Iran


Introduction: Content Based Image Retrieval (CBIR) is a method of image searching and retrieval in a 
database. In medical applications, CBIR is a tool used by physicians to compare the previous and current 
medical images associated with patients pathological conditions. As the volume of pictorial information 
stored in medical image databases is in progress, efficient image indexing and retrieval is increasingly 
becoming a necessity. 
Materials and Methods: This paper presents a new content based radiographic image retrieval approach 
based on histogram of pattern orientations, namely pattern orientation histogram (POH). POH represents 
the  spatial  distribution  of  five  different  pattern  orientations:  vertical,  horizontal,  diagonal  down/left, 
diagonal down/right and non-orientation. In this method, a given image is first divided into image-blocks 
and  the  frequency  of  each  type  of  pattern  is  determined  in  each  image-block.  Then,  local  pattern 
histograms for each of these image-blocks are computed.  
Results: The method was compared to two well known texture-based image retrieval methods: Tamura 
and  Edge  Histogram  Descriptors  (EHD)  in  MPEG-7  standard.  Experimental  results  based  on  10000 
IRMA  radiography  image  dataset,  demonstrate  that  POH  provides  better  precision  and  recall  rates 
compared to Tamura and EHD. For some images, the recall and precision rates obtained by POH are, 
respectively, 48% and 18% better than the best of the two above mentioned methods.   
Discussion and Conclusion: Since we exploit the absolute location of the pattern in the image as well as 
its global composition, the proposed matching method can retrieve semantically similar medical images. 


Main Subjects