Multiple Features Based Two-stage Hybrid Classifier Ensembles for Subcellular Phenotype Images Classification
Title:
Multiple Features Based Two-stage Hybrid Classifier Ensembles for Subcellular Phenotype Images Classification
Author:
Bailing Zhang - China Tuan D. Pham - Australia
Appeared in:
International journal of biometrics and bioinformatics
Paging:
Volume 4 (2010) nr. 5 pages 176-193
Year:
2010
Contents:
Subcellular localization is a key functional characteristic of proteins. As an interesting ``bio-image informatics'' application, an automatic, reliable and efficient prediction system for protein subcellular localization can be used for establishing knowledge of the spatial distribution of proteins within living cells and permits to screen systems for drug discovery or for early diagnosis of a disease. In this paper, we propose a two-stage multiple classifier system to improve classification reliability by introducing rejection option. The system is built as a cascade of two classifier ensembles. The first ensemble consists of set of binary SVMs which generalizes to learn a general classification rule and the second ensemble, which also include three distinct classifiers, focus on the exceptions rejected by the rule. A new way to induce diversity for the classifier ensembles is proposed by designing classifiers that are based on descriptions of different feature patterns. In addition to the Subcellular Location Features (SLF) generally adopted in earlier researches, three well-known texture feature descriptions have been applied to cell phenotype images, which are the local binary patterns (LBP), Gabor filtering and Gray Level Coocurrence Matrix (GLCM). The different texture feature sets can provide sufficient diversity among base classifiers, which is known as a necessary condition for improvement in ensemble performance. Using the public benchmark 2D HeLa cell images, a high classification accuracy 96% is obtained with rejection rate $21\%$ from the proposed system by taking advantages of the complementary strengths of feature construction and majority-voting based classifiers' decision fusions.
Journal description