Morphological Cell Classification under Weak Supervision: A Learning from Label Proportions Approach

Abstract

Abstract Classification of cells is an important field for biology and pathology research, and there have been many effective models integrating it with machine learning techniques. In the supervised learning scenario, annotations on cells are crucial for cell classification, and they can be obtained by cell image analysis and biological staining. However, due to restriction on cost, time efficiency and limitations of certain microscopes, obtaining annotations of all individual cells for the training of fully supervised machine learning models is always more challenging. For this reason, it’s highly necessary to investigate the feasibility of applying machine learning model on classification of cells with their morphological features, but only proportion labels for groups of cells are known. The models developed and tested in this thesis are from Learning from Label Proportions (LLP), a sub-domain of weakly supervised learning. LLP models approach data in bags instead of instances, aiming at prediction on instance level labels but assuming only proportions of each class in each bag are known. We developed several LLP based machine learning models specifically for tabular data containing morphological features of cells. We tested and compared these models, examining technique details and bringing future research directions. Our findings indicate that LLP models can achieve competitive performance on morphological cell classification with only proportion labels known, bringing values on further research in biology by reducing the need for comprehensive cell labeling.