Colon Cancer Prediction with Transfer Learning and K-Means Clustering

Abstract

The automatic diagnosis of colon cancer is important for patients and their prognosis through the analysis of histopathological images. Traditional feature extraction methods extract low-level image features, and prior knowledge is required to choose meaningful features, which may be modified significantly by humans. Hence, unsupervised and supervised deep convolutions neural networks were utilized to analyze histopathological images of colon cancer. To overcome the impact of the unbalanced histopathology images in sub-classes, the balanced sub-classes are turned right and left, up and down, and rotated counter clockwise by 90$^\circ$ and 180$^\circ$. The proposed experimental findings for supervised histopathological image classification of colon cancer demonstrate that $Inception\_V3$ and Inception $ResNet\_V2$ outperform current algorithms. These findings suggest that the $Inception\_ResNet\_V2$ network is superior deep learning architecture for analyzing histopathological images for diagnosis colon cancers. As a result, in order to perform unsupervised image analysis, $Inception\_ResNet\_V2$ is utilized to extract features from colon cancer histopathology images. In addition, a new autoencoder network was created to convert the features collected by $Inception\_ResNet\_V2$ to a low-dimensional space for image clustering analysis. The test results demonstrate that the proposed autoencoder network outperforms the $Inception\_ResNet\_V2$ network in terms of clustering.