An empirical study and comparative analysis of medical image retrieval and classification techniques

Image retrieval and classification in medical domain are the two important aspects in decision making and automatic annotation of benign and malignant images. These processes improve the decision making during decease identification. Image classification is usually done by checking image visual or semantic content similarity. Image content may be represented by its low level visual features referring to mathematical attributes or high level description based semantic attributes. Image similarity is verified by using similarity or dissimilarity measures or distance metrics. As image attributes are wide in range, the similarity measure that worked well for one particular feature set may not show the similar performance for the other. Based on it this paper explored geometrical, statistical and cumulative dissimilarity measures viz. Manhattan, Euclidean, Chebyshev, Cosine, Chi-square, Kullback, Jeffrey, Kolmogorov, Earthmover's and Cramer distances and compared their effect in similar looking image retrieval process. Experiments were done on these distance metrics with respect to appearance based features, intensity features, texture features and shape features. Through the experimentation, certain conclusions were drawn on the performance of these distance metrics with different feature sets in classification and retrieval. Comparative analysis performed on IRMA CLEFmed 2007 and 2008 image data sets. Mean Average Precision and Average Recall Rates were computed in analyzing retrieval performance task.