| In modern medicine,cancer,as one of the "murderers" that kills human beings,has always attracted much attention.Pathological slices have always played an important role in the diagnosis,treatment and prognosis of cancer.Pathologists can assess the development of the disease accurately by observing the morphology,location and number of cells in the pathological slices.However,pathological images contain complex tissue environments and tissue cells with different shapes.Pathologists need to consume huge time and labor costs when analyzing and discriminating the cells in the images.Therefore,the realization of computer-aided detection of cells in pathological images has become an urgent problem to be solved.Based on the above urgent needs,many scholars have carried out research on pathological image analysis,and have achieved remarkable results.However,these results are based on a large number of learning samples.Deep learning is based on a large amount of research data,and to obtain a model with good performance requires sufficient labeled data.Compared with natural images,cells in pathological images are detection targets with dense distribution and smaller scale.Therefore,manual labeling of cells is a very difficult task,which requires not only a careful labeling process,but also professional medical knowledge.Therefore,fully labeled pathological image datasets are not easy to obtain.In view of the above problem of insufficient data,from the perspective of few-shot learning,this paper proposes learning methods from two aspects of data and model,so as to improve the performance of the model with a small amount of annotation information.The contents of this paper are as follows:1.Although fully labeled pathology images are difficult to obtain,it is not difficult to collect a large number of unlabeled pathological slices.Based on this reality,this paper adds a large number of unlabeled images to the learning process of the model,and uses the pretraining model to generate corresponding pseudo-labels for the unlabeled images,thereby expanding the training dataset and enabling the model to learn sufficient morphological features and distribution feature of cells.When generating pseudo-labels,this paper adopts two methods,adaptive threshold and cell count,to remove the background noise caused by the complex tissue environment in pathological images,so as to ensure that the pseudo-labels have high confidence.These two pseudo-labeling methods select pseudo-labels from different perspectives.The pseudo-labeling method based on adaptive threshold obtains an adaptive threshold by quantifying model uncertainty,which is used to remove noise factors in the background to avoid generating false pseudo-labels in background regions.The pseudo-labeling method based on cell count generates a small number of pseudo-labels in the region with the highest response value,ensuring the accuracy of pseudo-labels.In these two pseudo-labeling methods,both the adaptive threshold and the number of pseudo-labels can be properly adjusted according to the different cell distributions in unlabeled images,so as to obtain the most accurate pseudo-labels.At the same time,when pseudolabels are used for model retraining,data distillation is performed on the feature map of unlabeled images,sacrificing a small amount of low-confidence pseudo-labels,thereby ensuring the reliability of the data involved in model learning.2.Aiming at the problem that cells in pathological images are densely distributed and even overlap each other,which makes cell detection difficult,this paper proposes a multi-task learning-based cell detection model from the perspective of model improvement.In this method,two related learning tasks of cell detection and cell count are integrated into the end-to-end model.Cell detection and cell counting share the same feature extraction module,and feature sharing is achieved through multi-task learning.In the learning process,cell count as an auxiliary task can provide global features of cell distribution and cell number for the cell detection task,thereby constraining the number of cells in the feature map,alleviating the detection errors caused by overlapping cells,and helping the cell detection network to locate the cell center more accurately.In addition,multi-task learning not only improve the effect of cell detection,but also accelerate the convergence speed of the model through joint learning.In this paper,the above method are verified on three datasets with different staining of different tissues.The results show that the detection effect of the model with a large number of unlabeled images involved in training is significantly better than that of the model that only uses a small amount of labeled data for pretraining,and above methods have good applicability on all three datasets.In addition,this paper also compares the experimental results of single-task learning（only cell detection task）and multi-task learning.In most cases,the multi-task learning model outperforms single-task learning.Moreover,compared with other semi-supervised learning methods,pseudo-labeling in this paper also has superior performance. |
