Poster Presentation 2014 International Biophysics Congress

The influence of image denoising on granulopoietic cell recognition using echo state networks (#526)

Philipp Kainz 1 , Michael Mayrhofer-Reinhartshuber 1 , Harald Burgsteiner 2 , Martin Asslaber 3 , Helmut Ahammer 1
  1. Institute of Biophysics, Medical University of Graz, Graz, Austria
  2. Institute for eHealth, Graz University of Applied Sciences, Graz, Austria
  3. Institute of Pathology, Medical University of Graz, Graz, Austria

Quantitative assessment of cellularity in human bone marrow is an essential step in medical diagnostics. The maturation continuum of white blood cells in bone marrow (granulopoiesis) is reflected by a minimal inter-class distance and a high intra-class variance, both of which are exacerbating the cell classification and often lead to high inter-observer variability. Whole slide imaging of histopathological specimen is required for computer-assisted cellular density estimation, but objects are usually degraded by noise in the digitalization process. Histopathological image analysis demands robust methods for tissue classification.1,2  Machine learning algorithms3,2 learn from examples and generalize to unseen data, which is a known disadvantage of classic image processing methods.4 

In this work we examined the influence of three popular denoising techniques (i) median filter (MF), (ii) Gaussian filter (GF), and total variation (TV5 ) on the classification performance of an Echo State Network (ESN6 ). Ground truth images (n=169) containing single cells of two subsequent maturation stages in granulopoiesis were labelled by an experienced pathologist. The ESN has been trained on raw and denoised image patches using ridge regression in order to discriminate between promyelocytes (n=84) and myelocytes (n=85).

Performance was assessed in 40 independent experiments in terms of overall mean accuracy (±SD): MF (0.824±0.009), GF (0.833±0.009), TV (0.820±0.008), raw (0.822±0.010). A Kruskal-Wallis-H test discovered statistically significant differences of the performance among some groups (p<0.001). All pairwise post-hoc Mann-Whitney-U tests revealed significant differences (Bonferroni-corrected p<0.008), except for MF vs. TV (p=0.055), MF vs. raw (p=0.525), and TV vs. raw (p=0.310).

With respect to overall mean accuracy, only Gaussian denoising influences (i.e. improves) the classification performance with statistical significance when compared to raw data. However, all data sets qualitatively yielded comparable results, which underlines the robustness of the ESN classifier against additive Gaussian noise.

  1. Gurcan, M. N., Boucheron, L. E., Can, A., Madabhushi, A., Rajpoot, N. M., Yener, B.: Histopathological image analysis: a review. IEEE Rev Biomed Eng, 2:147–171, 2009.
  2. Fuchs, T. J., Buhmann, J. M.: Computational pathology: Challenges and promises for tissue analysis. Comput Med Imag Graph, 35(7–8):515–530, 2011.
  3. Bishop, C. M.: Pattern Recognition and Machine Learning. New York: Springer, 2006.
  4. Gonzalez, R. C., Woods, R. E.: Digital image processing. Upper Saddle River, N.J.: Prentice Hall International, 2008.
  5. Chambolle, A., Caselles, V., Cremers, D., Novaga, M., Pock, T.: An introduction to total variation for image analysis. In Theoretical Foundations and Numerical Methods for Sparse Recovery. De Gruyter, pp. 263-340, 2010.
  6. Jäger, H.: The "echo state" approach to analysing and training recurrent neural networks - with an erratum note. GMD Report 148, 2001.