Patch Based Classification of Cell Painted ER and Cytoplasm using Block Intensity Gradient Pattern and Multilayer Perceptron

Differentiating subcellular structures in microscopic images is a difficult task due to their high similarity in visual appearance. In this work, an attempt has been made to classify Endoplasmic Reticulum (ER) and cytoplasm using Block Intensity Gradient approach and multilayer perceptron. For this, Cell Painted public dataset from Broad Bioimage Benchmark collection is considered. In an image patch small squared regions called blocks at multiple scales are selected. Horizontal and vertical gradient feature vector is extracted from all block pairs at each scale and concatenating them to form the full image descriptor. Statistically significant features are selected based on the p value (< 0.05) and these discriminative features are fed to the multilayer perceptron for categorizing ER and cytoplasm. The performance of BIGD descriptor for accurate classification is evaluated for different patch sizes and block sizes. The results indicate that there is an increase in classification accuracy with increase in image patch size for a fixed block size. 13x13 patch size with 4x4 block size achieved 96 % of classification accuracy. Experimental results demonstrate that BIGD texture feature could be useful for accurate classification of cell organelles by characterizing their dissimilarity in the cell structure due to cytological and cytotoxic effects.