Classification of Protein Crystallisation Images Using Texture-Based Statistical Features

We report on the classification of protein crystallisation data via decision trees using N-fold cross validation. Protein crystallisation images were obtained over a period of time ranging from days to months. All the images taken of a single experiment are arranged according to the time they were obtained and aligned according to the position of the experimental droplet. Difference between consecutive images in a time course was determined and the background noise was removed using filtering techniques. Image analysis is performed on the area inside the droplet of each difference image to compute statistical texture features. Classification is carried out by testing and training with 1000 sample images and an accuracy rate of similar to 75% was achieved.