Multi-class cancer diagnosis on histopathological images with deep ensemble learning model

Cancer is one of the leading causes of death worldwide. As the cause of one in six deaths, cancer causes a financial and emotional burden on individuals and a socio-economic burden on societies. Early detection of cancer significantly increases the chances of survival. In this study, we propose a computer-aided cancer diagnosis method that detects cancer on histopathologic images using current image processing and deep learning techniques and is intended to help diagnostic decision making. A stacking-based ensemble learning model using transfer learning is developed for the proposed cancer diagnosis method. In the proposed model, DenseNet 201 and EfficientNet B7 are used as base learners, and a two-layer CNN architecture is used as a meta-learner. The proposed model is trained and tested on a large dataset using lung, colon, oral and breast cancer data. When the results of the experimental studies are analyzed, the proposed model achieved 99 % accuracy on lung cancer data, 100 % accuracy on colon cancer data, 100 % accuracy on lung and colon cancer data, 85 % accuracy on breast cancer data, 86 % accuracy on oral cancer data and 90 % accuracy on a dataset containing four different cancer types with nine classes. Considering these results, the performance of the proposed model was found to be satisfactorily high. Thus, it is concluded that the proposed model can be used in computer-aided decision-making systems to assist medical professionals.