摘要
Abstract
A semi-supervised learning framework integrating rotational invariance,contrastive learning,and adaptive hybrid thresholds,named rotational contrastive network(RoCoNet),is proposed to enhance the applicability of semi-supervised learning for medical cell datasets.Due to the unique sampling approach of cell datasets,input images often contain uncertain rotation angles,which render traditional convolution kernels ineffective in existing semi-supervised detectors.To address this challenge,rotational attention convolution is introduced,offering robustness to rotational transformations.Additionally,cross-feature contrastive loss is proposed to improve upon the contrastive loss used in supervised learning,tackling issues of poor classification performance caused by cell overlap and clustering.An adaptive hybrid threshold is also introduced to stabilize pseudo-label generation during early training.A global threshold,computed by using Gaussian mixture models(GMMs),is applied to refine the local threshold,which helps balance the quantity and quality of pseudo-labels.Experiments on the ThinPrep cytology test(TCT)dataset for cervical cytopathology show that RoCoNet achieves a mean average precision(mAP)of 31.6%with only 10%labeled data,outperforming the baseline method by 8.4%in mAP.关键词
半监督学习/旋转不变性/对比学习/目标检测/宫颈细胞Key words
semi-supervised learning/rotational invariance/contrastive learning/object detection/cervical cell分类
信息技术与安全科学
