Abstract: Accurate detection of abnormal cells is essential for early screening and precise diagnosis of cervical cancer. Despite the recent advances in deep learning-based methods for cervical cancer detection, their broad clinical applications are hindered by several technical challenges. On the one hand, gradually evolved abnormal cells are visually similar to normal cells. On the other hand, single cells and cell clusters exhibit significant appearance variations, overlooking those between normal and abnormal cells. In order to overcome these challenges, we propose a novel dual-branch multi-granularity hierarchical contrast and cross attention network, called DMCA-Net. Specifically, DMCA-Net utilizes dual branches to detect abnormal and normal cells, respectively. Meanwhile, an inter-cell pair-wise cross-attention (IPCA) is utilized to improve feature embedding learning. The IPCA regularizes the attention learning of abnormal cell features by treating normal cell features as distractors. In addition, DMCA-Net also adopts a multi-granularity hierarchical contrastive learning (MHCL) to enhance the classification ability. Our study indicates that MHCL alleviates the interference of intra-class appearance variations in cervical cell, effectively pulls apart the inter-class distance between different classes of cervical cells at different granularities. Extensive experiments on two publicly available datasets demonstrate that our DMCA-Net outperforms existing methods, achieving state-of-the-art (SOTA) results.
Keywords: Cervical abnormal cell detection; Medical image analysis; Attention mechanism; Contrastive learning; Hierarchical classification