Classification of Glandular Cells using a Pre-trained Convolutional Neural Network
Migyung Cho

Migyung Cho, College of Software Convergence, Tongmyong University, Sinseon-ro, Nam-gu, Busan, Korea.
Manuscript received on 17 August 2019 | Revised Manuscript received on 27 August 2019 | Manuscript Published on 16 September 2019 | PP: 55-59 | Volume-8 Issue-2S6 July 2019 | Retrieval Number: B10110782S619/2019©BEIESP | DOI: 10.35940/ijrte.B1011.0782S619
Open Access | Editorial and Publishing Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract: Tissues can be divided into glandular structures (cells) and non-glandular when we magnify them with a microscope. To classify the two types, we performed experiments using convolutional neural network. We cropped regions of glandular cells and non-glandular in 10 x magnification images. The size of cropped images was between 80 and 100 pixels in width and height. We prepared 932 glandular cells and 1000 non-glandular for the train and test. Of these, 1468 were used for learning and 532 were used for testing. We trained and tested the dataset using a slightly modified pre-trained VGG16. The inside of glandular cells consists of nucleus, lumen and cytoplasm. Normal glandular cells and abnormal glandular cells that we call tubular adenoma have different texture features. But both types of glandular cells have distinct boundaries and specific shapes. In the case of cancer, as the nucleus grows excessively, the boundaries of the glandular cells become unclear and disappear. We trained three types of glandular cells, which is normal, tubular adenoma, and cancer. Experimental results using the pre-trained VGG16 classification showed a high classification accuracy of 99.44%. Only three non-glandular out of the 532 test data were misclassified into glandular cells. The classification method presented in the paper can be used to eliminate false positives that produced by an automatic segmentation system for the pathology image. The performance of the segmentation can be improved by eliminating segmented objects that are false positives.
Keywords: Classification, Convolution Neural Network, Glandular Cells, Pathology Image, Segmentation.
Scope of the Article: Classification