Feature Selection and Analysis of Diffraction Images

3D morphological features of biological cells captured in diffraction images provide the basis to deeply study the biochemical processes that a cell undergoes during various activities. Although automatic classification of biological cells is highly desired, manual analysis of image data is often necessary to study the underlying morphological processes as in most of the cases due to complex cell structure. The primary objective of the experimental study presented in this paper is to select the most discriminating GLCM features, and deeply investigate the quantization and displacement factors of GLCM on the performance of two classifiers, SVM and LMT for diffraction images. A total of 20 GLCM textural features for 6 types of cultured cells (100 samples per cell) were computed in this study. Without any feature selection the maximum classification accuracy was 90.8%, 86.83% with 10FCV using SVM and LMT, respectively. An optimal subset of features was selected using 2 different approaches. In approach 1, feature selection was done using EFCS resulting in 8 GLCM textural features. In approach 2, feature selection was done using CFS resulting in 8 different features. Using features from approach 1, the classification accuracy of SVM was increased from 90.8% to 91.16% while it remained the same in the case of approach 2. Also, features selected using approach 1 yielded better classification rates than features selected using approach 2 for both classifiers. The findings in this experimental study conclude that a set of 8 GLCM features selected using EFCS approach with a 6-bit quantization scheme along with a displacement of d=2 are the best for the classification of diffraction images.