Document Citation

Title:
A Classifier Ensemble Based On Performance Level Estimation.

Author(s):
Wang W, Zhu Y, Huang X,Lopresti D, Xue Z, Long R, Antani S, Thoma G.

Institution(s):
1) Department of Computer Science and Engineering, Lehigh University, Bethlehem, PA 18015
2) Communications Engineering Branch, National Library of Medicine, MD 20894

Source:
2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro. Boston, MA. 2009:342-45.

Abstract:
In this paper, we introduce a new classifier ensemble approach, applied to tissue segmentation in optical images of the uterine cervix. Ensemble methods combine the predictions of a set of diverse classifiers. The main contribution of our approach is an effective way of combination based on each classifier’s performance level—namely, the sensitivity p and specificity q, which also produces an optimal estimate of the true segmentation. In comparison with previous work [1] that utilizes the STAPLE algorithm [2] for performance level based combination, this work achieves multiple-observer segmentation in a Bayesian decision framework using the maximum a posterior (MAP) principle, considering each classifier as an observer. In our experiments, we applied our method and several other popular ensemble methods to the problem of detecting Acetowhite regions in cervical images. On 100 images, the overall performance of the proposed method is better than: (i) an overall classifier learned using the entire training set, (ii) average voting ensemble, (iii) ensemble based on the STAPLE algorithm; it is comparable to that of majority voting and that of the (manually picked) best-performing individual classifier in the ensemble set.

Publication Type: CONFERENCE

More about this article:

Full Text (PDF)