SKC-Net: A Standard-Cell Key-Feature Contrastive-Learning network for cervical abnormal cell detection

Deep learning-based detection of abnormal cells in cervical ThinPrep Cytology Test (TCT) images represents a pivotal advancement in intelligent cervical cancer screening systems. In this context, the effective utilization of standard cell features in images is paramount. On one hand, the TBS diagnostic criteria delineate the differentiation based on feature comparisons between normal and abnormal cells. On the other hand, significant inter-sample variations can lead to confusion during the learning process in conventional object detection networks. To mimic the diagnostic reasoning of professional pathologists, this paper introduces the Standard-Cell Key-Feature Contrastive-Learning Network (SKC-Net) for detecting cervical abnormal cells. The network initially extracts key features from both standard and abnormal cells and effectively aligns them using consistency-guided contrastive loss. To better distinguish the unique characteristics of standard and abnormal cells, we have designed a Multi-Color-Domain Convolution Module (MCD) and an Adaptive Contour Approximation Module (ACA), which are based on TBS diagnostic criteria—focusing respectively on the preliminary extraction of color and shape features. Furthermore, to mitigate information loss during deep feature learning, we propose a Dual-Branch Differentiated Convolution Module (DBD) aimed at enhancing the model’s learning capacity and memory retention. Finally, to guide the model’s learning process by leveraging standard cell key features, we introduce the Cell Instance-Level Contrastive Supervision Module (CISM); this enhancement significantly boosts the overall model performance. Experimental results demonstrate that our method improves the mean Average Precision (mAP50) by 3.1% and 3.0% on the CLBC and CRIC datasets, respectively, achieving scores of 65.2% and 71.8%. These findings underscore the effectiveness of our proposed approach.