CELL IMAGE CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Medical image processing is a field of great interest, and improvements on this field have made possible better and faster diagnosing of sick organs or tissues. This study’s main focus is the classification of healthy and unhealthy cells. In this study we have discussed and compared the behavior of the LeNet network in different given conditions of the network and dataset. There have been considered three different data splitting: the first one having two classes and a dataset of 9 332 images, the second one having two classes and a dataset of 20 102 cell images and the third data split having three classes and 12 520 images. All these cases were trained and tested in similar and different network conditions and preprocessing methods to be able to evaluate which one of them performs better with the available datasets. The main preprocessing methods used are unsharped masking, median filter and highpass filter. Moreover the models were compared in pairs using AUC and ROC curve, in order to distinguish even the slightest changes and improvements on models.