Advances in artificial intelligence (AI) have powered many new technologies in the realms of natural language processing (e.g. Siri or Alexa) and computer vision (object detection and classification in videos and images, e.g. self-driving cars). I am particularly interested in the use of such machine learning systems for the analysis of medical images. While the field of radiology is working with digital images since a long time, the related discipline of pathology is only recently undergoing the digital transformation. Here, we are dealing with patient specimens in the form of tissue or cells for the study of disease. Recently, whole slide image scanners are available, that enable the rapid digitization of traditional stained glass slides into digital gigapixel images. The availability of digital images coupled with new and powerful AI methods gives rise to computational pathology which is concerned with the development of novel decision support systems as a way to further personalized medicine efforts and enable differential diagnosis systems for early detection and characterization of diseases such as cancer.
For my Gateway Fellowship project I could win the support of Prof. Jun Sakuma from the University of Tsukuba in Japan. His Machine Learning and Data Mining lab is very much theory focused while I come from a more practical bioinformatics background. The study and subsequent integration of foundational new ideas into established state-of-the-art approaches is a way to synthesize novel and powerful applications. As such, I want to investigate the potential of very scalable but inefficient transformer architectures that currently dominate many other areas of computer vision research but could not yet be successfully applied to medical images due to their sheer size. In Japan, there is the opportunity to work on a distributed network of hospitals to enable federated learning, which would split the computational load across many instances, thus giving me the optimal conditions for this project.