The analysis of histopathological images of colorectal cancer (CRC) using machine learning methods

Abstract

With this drastic increase in cancer rates and deaths because of cancer, it is essential to find new and improved ways to prevent it as soon as possible (at early stages), especially colorectal cancer, which is ranked second in the cancers that cause death. Nowadays, it is occurring more and more among young people, often at an advanced stage. As it does not have specific symptoms, colorectal cancer is difficult to diagnose. Also, it is crucial to differentiate the benign lesions from the malignancies. Thus, a decision-support system was developed in this study to support this purpose. For this aim, a dataset of histopathological images with benign and malignant cases was utilized to generate features (68 in total), and statistical analysis was performed to find the features that can differentiate the categories. Furthermore, these features were used as input for creating a machine-learning system. Among the classifiers that we used were the K-Nearest Neighbors (KNN), Classification Trees (CART), and Random Forest. Recursive Feature Elimination (RFE) was employed for feature reduction, and the evaluation was conducted using Bootstrap and K-Fold cross-validation, focusing on accuracy, precision, and recall metrics. Moreover, further investigation was made to differentiate the two categories by employing some Convolutional Neural Networks (CNNs) utilizing pre-trained models, such as Vgg16, MobileNetV2, ResNet50V2, InceptionV3, and EfficientNetB0. The research proceeded by splitting the two categories into five (healthy, benign, grades I, II, and III) and testing every combination in pairs of two. The same steps, as described for the two, were followed for the five categories. The top five features that were observed to distinguish the two classes were the skewness, energy range, correlation mean, LBP6, and Gabor energy range. Also, the RF classifier using the bootstrap evaluation method satisfactorily differentiated benign from malignant with an accuracy of 90.20%, 87.63% sensitivity, and 91.99% specificity. Lastly, the RES50 model showed the best accuracy of 90.92% -for 100 epochs. Our system also performed adequately for the five categories despite the reduced size of the data per class.