Optimized deep learning architectures for the classification of colorectal cancer diagnosis using whole slide images

Background The overwhelming number of cancer cases around the world has expressed a critical need for an automated diagnostic tool to assist pathologists in efficiently handling these cases. Colorectal cancer is one of the most common diseases in the world, increasing yearly. The integration of deep learning architectures in digital pathology has shown promising potential as a supportive tool for assisting pathologists in the diagnosis of cancerous tissues. However, the lack of histopathological image datasets of colon cancer impedes the precise evaluation of deep learning diagnosis techniques. Methods This study proposes an ensemble model, combining EfficientNetv2 and DenseNet architectures, for the binary classification of colorectal cancer from whole slide images. The framework utilizes a new custom dataset containing histopathological images of colorectal cancer cases divided into benign and malignant classes, collected from Bahrain Defence Force-Royal Medical Services-King Hamad University Hospital in the Kingdom of Bahrain. The dataset comprises a total of 4,694 images, extracted from 227 whole slide images of colorectal cancer patients. However, due to limited computational resources, only 2,000 images were utilized in this study. Results The proposed model achieved a commendable accuracy of 98%, a perfect precision of 100% and a recall of 96.30%, displaying a high generalization ability and robustness. Furthermore, a comparative analysis was performed, which showed that the proposed model outperformed several state-of-the-art architectures.