Abstract
With availability of huge amounts of labeled data, deep learning has achievedunprecedented success in various object detection tasks. However, large-scaleannotations for medical images are extremely challenging to be acquired due tothe high demand of labour and expertise. To address this difficult issue, inthis paper we propose a novel semi-supervised deep metric learning method toeffectively leverage both labeled and unlabeled data with application tocervical cancer cell detection. Different from previous methods, our modellearns an embedding metric space and conducts dual alignment of semanticfeatures on both the proposal and prototype levels. First, on the proposallevel, we generate pseudo labels for the unlabeled data to align the proposalfeatures with learnable class proxies derived from the labeled data.Furthermore, we align the prototypes generated from each mini-batch of labeledand unlabeled data to alleviate the influence of possibly noisy pseudo labels.Moreover, we adopt a memory bank to store the labeled prototypes and hencesignificantly enrich the metric learning information from larger batches. Tocomprehensively validate the method, we construct a large-scale dataset forsemi-supervised cervical cancer cell detection for the first time, consistingof 240,860 cervical cell images in total. Extensive experiments show ourproposed method outperforms other state-of-the-art semi-supervised approachesconsistently, demonstrating efficacy of deep semi-supervised metric learningwith dual alignment on improving cervical cancer cell detection performance.