Early Detection of Colon Cancer on Histopathology images

Abstract

Colorectal cancer is a type of cancer that develops in the colon or rectum. It typically starts as small noncancerous clumps of cells called polyps that can become cancerous over time. In Colorectal cancer, Histopathology images play a critical role in the diagnosis and treatment, it is a diagnostic technique that involves the examination of tissues under a microscope to identify abnormalities in the cells and tissue architecture. In recent years, advances in digital pathology have made it possible to analyze histopathology images using artificial intelligence and machine learning algorithms. Deep learning, in conjunction with convolutional neural networks (CNNs), has emerged as a powerful and transformative approach in recent years, particularly in the medical field. It has made significant advancements across various domains, revolutionizing the way we analyze and interpret medical data. In this thesis, our primary focus is on the classification of multiple tissue types within colorectal cancer (CRC) samples using a widely used database known as CRC-HE-VAL-7K, which comprises 7,000 images. For our CNN models, we explore and evaluate three distinct techniques. Firstly, we employ neural network training from scratch, allowing the network to learn directly from the given data. Secondly, we utilize transfer learning by leveraging the pretrained VGG19 model, which has demonstrated exceptional performance in image recognition tasks. Lastly, we propose an ensemble CNN approach, which combines the strengths of neural network training from scratch, VGG19, ResNet50, and Inceptionv3 models. Furthermore, to achieve better results and improve overall accuracy, we incorporate two ensemble methods: the Average method and the weighted averaging method. These methods enable us to combine the predictions from multiple models in a systematic manner, taking into account their respective performance and assigning appropriate weights to the individual models