Image Processing Reference
In-Depth Information
Automatic classification of
protein crystal images
Madhav Sigdel 1 ; Madhu S. Sigdel 1 ; İmren Dinç 1 ; Semih Dinç 1 ; Marc L. Pusey 2 ; Ramazan S. Aygün 1
1 DataMedia Re-
search Lab, Computer Science Department, University of Alabama Huntsville, Huntsville, AL, USA
2 iXpressGenes Inc., Huntsville, AL, USA
This work introduces our method for automatic classification of crystallization trial images according to
the types of protein crystals present in the images. The images are classified into four categories: needles,
small crystals, large crystals, and other crystals. Because protein crystals are characterized by some geo-
metric shapes, we focus on extracting geometric features from the images. Our image feature extraction
includes extraction of blob features from multiple binary images, extraction of edge related features from
Canny edge image, and extraction of line features using Hough line transform. For the decision mod-
el, we propose applying random forest classifier. We performed our experiments on 212 expert labeled
images with different classifiers and tested our results using 10-fold cross validation. The proposed clas-
siication technique produces a reasonable performance for protein crystallization image classification
The overall accuracy using random forest is 78%.
Protein crystallization
Crystal classiication
Blob features
Edge features
This research was supported by National Institutes of Health (GM090453) grant.
1 Introduction
Protein crystallization is the process for formation of protein crystals. Success of protein crystal-
lization is dependent on several factors such as protein concentration, type of precipitant, crys-
tallization methods, etc. Therefore, thousands of crystallization trials with different crystalliz-
Search WWH ::

Custom Search