Performance of medical image retrieval techniques: A comparative analysis

Medical Image retrieval and similar image classification are the two important aspects in decision making and automatic medical image annotation. These approaches help the doctors and radiologists in their decision making during decease identification and decision making. Image classification is usually done by checking image visual content or semantic content similarity. Image content may be represented by its visual features referring to mathematical attributes or high level description based semantic attributes. Content similarity is done by using similarity and dissimilarity measures or distance metrics. As image attributes are wide in range, the similarity measure worked well for one feature set may not show the similar performance for other. For this reason in this paper we explored geometrical and statistical dissimilarity measures viz. Manhattan, Cosine, Chi-square and Cramer distances and their effect with respect to image intensity feature set and wavelet based texture feature set. Through experimentation, we have drawn certain conclusions on the performance of these distance metrics in classification and retrieval of IRMA CLEFmed 2007 and 2008 image sets. Mean Average Precision and Average Recall Rates were used in analyzing retrieval performance for analyzing the medical image retrieval and classification task.