Research On Cell Detection In Pathological Images Based On Transfer Learning

Pathological diagnosis is the gold standard of most diseases,and the information provided by pathological tests has become the key proof for many diseases.Cell detection of pathological images is of great significance for cancer diagnosis,grading,and prognosis.Although supervised methods have shown great performance on cell detection for pathological images,high cost of manual pixel-level annotation makes it hard access.In addition,cells in different staining have diversity of visual characteristics.Therefore,it is a great requirement to develop cross-modal for detecting cells in pathological images.This article aims to apply transfer learning method to tackle the problem of insufficient pixel-level annotations in pathological images.Currently,pathological image data sets with a large number of annotations are very rare,and some data sets are not fully labeled.For those pathological images without any annotations,an unsupervised method based transfer learning for cell detection is proposed.Meanwhile,we also develop a semi-supervised transfer learning method for those with few annotations.Focusing on above two issues,this article includes efforts which are as follows:1.In this article,we design an unsupervised pathological image cell detection method based on transfer learning,which is an end-to-end adversarial learning model with unsupervised domain adaptation for cell detection.The mutual conversion module of different stained pathological images and the cell detection module are combined into an end-to-end model to achieve mutual restriction of accuracy.In addition,we propose a cross-domain consistency loss method,which can simultaneously modify the results of image transformation and cell detection.The results of comparative experiments demonstrate that our proposed method achieves the optimal performance among all compared with other methods.2.We designed a semi-supervised pathological image cell detection method based on transfer learning,which is an application of semi-supervised semantic segmentation based on cross-consistency in domain adaptation.Consistency training is a powerful semi-supervised learning framework using unlabeled data under the assumption of clustering,in which different perturbations applied to the encoder output force the prediction to be invariant.Specifically,the available annotation data is used to train the shared encoder and the main decoder in a supervised manner.In order to utilize the unlabeled data,we maintain consistency between the prediction of the main decoder and the prediction of the auxiliary decoder,and use the different interference versions output by the encoder as input,thereby improving the representation of the encoder.Experimental results demonstrate that the performance of our proposed method is better than that of a fully supervised method with a small amount of labeled data.