Detecting colon cancer using deep learning on segmented histopathological images

This study is based on color human colon tissue images, shot with a microscope mounted camera and dyed with the Hematoxylin and Eosin dye technique. The study is composed of two stages. The first stage led to the proposal of the augmented k-means clustering algorithm for the segmentation of cells and cell nuclei on colon tumor images. In the proposed augmented k-means clustering algorithm, the starting cluster centers were identified with the assignment of the values obtained by dividing the color range by the number of clusters. The proposed algorithm was then compared with the weighted k-means clustering algorithm. During the comparison, all processes were repeated 3 times under the same set of conditions, with a view to analyzing the performance and stability of the algorithms. The experiments revealed that the augmented k-means clustering algorithm reduced the iteration count and consequently the process time. Moreover, the use of the proposed approach instead of the random positioning of the initial cluster centers was observed to lead to higher stability. In addition, the similarity of the images obtained through segmentation with the original ones was assessed using the histogram-based similarity algorithm. The assessment found that the use of the augmented k-means clustering algorithm produced images which are more similar to the original images. In the second stage, on the other hand, two models of convolutional neural networks –AlexNet and GoogLeNet– were employed, culminating in a new approach towards the classification of colon cancer. In this context, for the purpose of training the convolutional neural networks, the images segmented into color clusters through the segmentation method proposed in the first stage, instead of the original-raw images, were employed. For this purpose, 20 datasets of images which are distinct in terms of their structure and characteristics were derived from the larger datasets comprised of the original-raw images, as well as the segmented images. The datasets thus produced were then employed with the AlexNet and GoogLeNet convolutional neural network models, to train them and for testing. The test results provided the input for the development of confusion matrices, drawing of ROC curves, and the calculation of AUC values. The results show that the AlexNet model trained through the use of the segmented images registered a 2% to 23% increase in the model performance, whereas the GoogLeNet model thus trained had a 2% to 27% increase in the model performance. Furthermore, the proposed approach was found to lead to higher performance with the datasets in which the data was not homogenous.