Ensemble CNN for colon cancer detection using histopathological image

Colon cancer remains one of the leading causes of cancer-related mortality worldwide, with delayed detection significantly impacting survival rates. The time-consuming nature of traditional histopathological analysis contributes to these diagnostic delays. This study leverages advances in artificial intelligence to improve both the speed and accuracy of colon cancer detection through automated analysis of tissue sample images. We propose a weighted ensemble model comprising four convolutional neural networks namely RegNet X, RegNet Y, Swin B, and VGG11, that enhances prediction accuracy by combining outputs from base models. The prime focus of the research is to devise a recall prioritized ensemble model capable of detecting various types of colon cancers, while also examining the possibility of the model to generalize beyond the training domain. The ensemble was trained using a dataset of 10,000 histopathological images of colon tissues containing both benign and malignant samples. We introduce a novel preprocessing pipeline incorporating histogram equalization and contrast stretching for image enhancement with a motive to improve diagnostic accuracy. Notably, the ensemble model demonstrated superior recall metrics compared to traditional ML approaches—a critical measurement in healthcare applications. The ensemble model proposed has a test accuracy of 99% and a recall of almost 100%. Further, the proposed model is also tested for lung cancer to validate its generalizability. The experimental results obtained establish the validation of proposed model in terms of generalizability, demonstrating strong performance.