Weakly Supervised Classification using Multi-Level Instance-Aware Optimization on Cervical Cytologic Image

Abstract

The pathological images of liquid-based cytology are widely used in cervical cancer screening, and its large resolution has always limited the efficiency of diagnosis. Weakly supervised learning is an efficient method for computer-aided diagnosis. However, its performance may also be limited by the rough annotation. Therefore, we propose an optimized multi-instance classification framework to learn more reliable representation from multi-level instance awareness. We first introduce deep self-attention modules following various layers of the instance-level encoder, which promotes the model to learn the relationship between instances. Then we cluster the instance features in each bag to strengthen distinguishability. In addition, we propose an adaptive instance mask strategy to facilitate the learning of relevant features from suspicious samples with weak attention. Our method performs a significant improvement by comparing with competitors, and attention visualization also reveals its effectiveness.