Oops, you're using an old version of your browser so some of the features on this page may not be displaying properly.

MINIMAL Requirements: Google Chrome 24+Mozilla Firefox 20+Internet Explorer 11Opera 15–18Apple Safari 7SeaMonkey 2.15-2.23

Poster display

4141 - Qualitative and quantitative differences in a set of criteria for histological grading of colorectal cancer


10 Oct 2016


Poster display


Taras Savchyn


Annals of Oncology (2016) 27 (6): 545-551. 10.1093/annonc/mdw393


T. Savchyn1, S.A. Antoniuk2, A.N. Grabovoy2

Author affiliations

  • 1 Esc “institute Of Biology”, Taras Shevchenko National University of Kyiv, 04655 - Kyiv/UA
  • 2 Department Of Pathologic Anatomy, National Cancer Institute of the MPH Ukraine, Kiev/UA


Abstract 4141


Colorectal cancer grading still remains semiquantitative and subjective. The aim of this study was to estimate the differences between G1, G2 and G3 tumor grades using quantitative methods and build objective tests for tumor grading.


Samples were collected from 114 patients with colorectal cancer. The tissue slices were stained with Einarson's gallocyanin chrome alum stain for the NA content detection. Immunohistochemical reactions were conducted using monoclonal antihuman Ki-67, Bcl-2 and p53 antibodies. Silver nitrate impregnation was used for nucleolus organizer regions (NORs) detection. Regression analysis was conducted using logistic model. Cox and Snell's R2 (R2CS), Nagelkerke's R2 (R2N), Bayesian information criterion (BIC) and ROC curves (AUC values) were used for assessing the fitting ability of obtained models.


Two strategies for multiclass classification were used. One-vs.-one (OvO) compares grades pairwise and then builds binary classifier for each grade. One-vs.-all (OvA) strategy involves training classifier based on comparing the grade to the set of others. Logistic regression model was selected as binary classifier. As a result large varieties of models were built including different set of predictors. Models with the best accuracy were selected for each strategy (Table). G2 vs. G3 includes such predictors as tumor expression rate of Bcl-2 and p53. And these predictors are included in G3 vs. All. Pointing that the main effect induced by G2 and G3. But G2 vs. G3 has better accuracy (AUC = 0.8). G1 vs. G2 includes number of NORs total volume of NORs and average DNA content in the cells. G1 vs. All includes the same predictors too but shows better accuracy (AUC = 0.74).

χ2 df p R2CS R2N BIC AUC
OvO G1 vs. G2 15.38 3 0.002 0.15 0.22 124 0.72
G2 vs. G3 14.82 4 0.005 0.17 0.29 91 0.80
G3 vs. G1 19.52 3 0.0002 0.39 0.54 53 0.89
OvA G1 vs. All 18.61 3 0.0003 0.16 0.24 130 0.74
G2 vs. All 7.23 2 0.03 0.07 0.09 154 0.64
G3 vs. All 13.91 4 0.008 0.12 0.23 103 0.78


Therefore, models from different strategies can be combined into a novel one. In our case G1 vs. All and G2 vs. G3 may be used together for colorectal tumor grading in an hierarchical manner. First off G1 separates from group of G2 and G3 with the following G2 and G3 separation.

Clinical trial identification

Legal entity responsible for the study

National Cancer Institute, Department of Pathologic Anatomy; 33/43, Lomonosova str., Kyiv, 03022, Ukraine


Ministry of Healthcare of Ukraine


All authors have declared no conflicts of interest.

This site uses cookies. Some of these cookies are essential, while others help us improve your experience by providing insights into how the site is being used.

For more detailed information on the cookies we use, please check our Privacy Policy.

Customise settings