Research On Pathological Image Cell Detection Under Limited Labeled Data

Pathological image examination is hailed as the gold standard for diagnosing cancer,prognosis and guiding treatment,and it is also a key link between diagnosis and treatment.It makes a diagnosis by observing the structure of biopsied material and the characteristics of cytopathic changes.The results are more authoritative than other diagnostic methods.It is currently the most important and reliable cancer diagnosis method.With the continuous development of digital imaging equipment and pathological section making technology,pathologists can use advanced imaging equipment to scan biopsied samples and observe the whole slide images on computer.However,the resolution of whole slide images is extremely high,and the content of the images is complex,covering millions of cells with diverse shapes.It is obviously time-consuming and labor-intensive to rely solely on pathologists to analyze such a complex task with the naked eye,and there is an urgent need for automated technology to assist doctors in analyzing the content of pathological images.Existing related technologies based on deep learning are based on a large amount of high-quality labeled data,and the cost of labeling pathological image data is high,and usually only a limited amount of labeled data can be obtained for model training.Therefore,there is an urgent need to carry out research on pathological image cell detection under limited annotations.This article focuses on the detection of pathological image cells under limited annotation,and has achieved the following research results:First of all,a pathological image cell detection network CRDNet based on classification reinforcement branch is proposed.There are a large number of normal cells and abnormal cells with very similar appearance and shape in pathological images.It is difficult for conventional detection networks to distinguish between the two,which is likely to cause a large number of missed detections and false positives.This problem is even more prominent under the condition of limited labeled data.CRDNet is composed of a basic detection network and additional designed classification reinforcement branch.It strengthens classification capabilities from two aspects: rational use of context information and enhancement of the network’s feature representation,effectively distinguishing normal cells from abnormal cells.Experiments on the signet ring cell dataset prove that it can successfully suppress the missed detection and false positives results,and effectively improve the detection effect of pathological image cells under limited annotation.Secondly,a data distillation-based semi-supervised pathological image cell detection framework DDCDF is proposed.Since it is difficult to obtain sufficient high-quality labeled data in the field of pathological images,DDCDF focuses on how to use massive unlabeled data to help optimize the detector.DDCDF includes an improved data distillation mechanism and a similarity augmentation strategy.The former makes the generated pseudo-labels with high credibility by continuously maintaining high-quality supervision information during the training process,and the latter enriches the training set information by synthesizing new data.Experiments on the signet ring cell dataset prove that DDCDF effectively uses unlabeled data to improve the performance,which is a great improvement compared to mainstream methods.