Self-supervised learning architectures as the basis for foundation models in histopathology : A comparative study

Abstract

Artificial intelligence driven decision support systems will revolutionize clinical pathology in the near future. Machine learning has been shown to produce models that match or even exceed the performance of clinical experts. The demand for these kinds of systems is also high with the continuously increasing number of histopathological samples to be analyzed. Training an accurate model in a supervised setting requires a large amount of pre-labeled training data which imposes practical limitations on the usability of this kind of approach. An alternative approach is to use self-supervised learning, where large amounts of unannotated data can be leveraged to learn useful representations that can be used as a basis for the supervised training.

In this paper, we study how much training data is actually required to train an accurate state-of-the-art deep learning ensemble model for prostate and breast cancer detection from hematoxylin and eosin stained whole slide images using supervised learning. Furthermore, we explore the use of multiple self-supervised training approaches and study if the number of training samples can be reduced without a penalty to the performance of the trained models. The results provide strong evidence in support of using self-supervised pre-training to make cancer detection machine learning models more practical and data-efficient to train.