Human-in-the-Loop Deep Learning for Oral Cancer Screening: An Active Learning Approach

Despite early detection being widely recognized as a key strategy to reduce the global burden of oral cancer, there are currently no widely adopted oral cancer screening programs. Following the advancements in artificial intelligence and deep learning over the last decade, new possibilities in medical imaging have emerged, and a deep learning-based pipeline for oral cancer screening on cytological whole-slide images has been proposed. This thesis investigates whether an active learning-based human-in-the-loop approach can improve the performance of deep learning-based nucleus detection in the proposed pipeline in a label-efficient manner. By considering both theoretical and practical aspects of label efficiency, we thoroughly examine three active learning methods and establish a practical workflow for interactive annotation with large-scale applications such as mass screening in mind. Experimental results demonstrate that active learning can significantly improve the label efficiency of nucleus detection compared to a baseline method of random sampling. In combination with the developed infrastructure, this thesis highlights the potential impact of active learning-based human-in-the-loop solutions for large-scale clinical applications.