Efficient Strategy for Building Colorectal Cancer Classification CAD System Using Weakly-Annotated Whole Slide Images

Colorectal cancer (CRC) is one of the most deadly diseases nowadays. With the advancement of deep learning, it has become feasible to build CAD (computer-aided diagnosis) systems that help doctors in the process of disease diagnosis and prognosis. In this research, we study the CRC classification problem using weakly-annotated whole slide images (WSIs) based on the multiple instance learning (MIL) approach. We concentrate on applying efficient preprocessing steps that output a computational power-efficient representation of the data; and performing a set of experiments to determine the best model structures and hyperparameters suiting the problem being studied, which serves as a preliminary step toward improving the learning algorithm. Our research proves that the correct choice of these factors can potentially improve the learning process and the model performance. The results we achieved demonstrate significant superiority over the baseline work that leveraged the weakly-annotated slides, where we achieved an 89.58% accuracy compared to 84.17%. Our results even outperformed the original work evaluation results obtained by applying a pre-training step using a set of strongly-annotated slides.