LESS: Label-efficient multi-scale learning for cytological whole slide image screening

Beidi Zhao, Wenlong Deng, Zi Han henry Li, Chen Zhou, Zuhua Gao, Gang Wang, Xiaoxiao Li

Abstract: Highlights•Propose a novel multiple instance learning (MIL) framework for label-efficient cytological whole slide image (WSI) screening with slide-level label only.•Investigate intrinsic properties of cytology WSIs and be the first to adapt variational positive-unlabeled learning in WSI analysis for patch feature extraction.•Provide a multi-scale cross-attention-based Transformer aggregation scheme to capture complementary information of small and large patches.•Perform evaluation and interpretation of our pipeline on multiple cytology datasets with extensive comparison with state-of-the-art baselines.