Study On Automatic Analysis Methods For Digital Pathology Images Based On Semi-Supervised Deep Learning

Pathology images are the gold standard for most cancer clinical diagnosis,and automated analysis of pathology images can help pathologists reduce the workload as well as improve the efficiency and the accuracy of cancer diagnosis.Recently,automatic pathology image analysis methods based on deep learning have achieved success,but their prediction performance heavily rely on using a large number of labeled pathology images.Since pathology images data can only be annotated by professional pathologists and the annotation process is very complicated,it is important to develop automatic pathology image analysis methods based on semi-supervised learning,so that they can make full use of the hidden information in unlabeled data to improve model performance with limited labeled data,which have important research significance and application prospects.However,the existing researches relate to semisupervised pathology image analysis are still relatively limited,and there are many issues need to be further explored in terms of technical tools and performance improvement.In this study,we conduct an in-depth study and propose various novel semi-supervised deep learning methods for image classification and nuclei detection tasks in analysis of pathology images.The main contributions of this study are as follows:(1)In order to improve the classification performance of pathology images with a small amount of labeled data,a semi-supervised pathology image classification method called Semi-HIC is proposed.Specifically,Semi-HIC method first designs a convolutional neural network based on cascaded Inception blocks(CIB)to extract rich and discriminative features from pathology images.Subsequently,based on the associative learning,association cycle consistency(ACC)loss and maximal conditional association(MCA)loss functions are introduced to train the neural network.The aforementioned losses effectively address the wide intra-class variability and inter-class similarity problems in pathology images by suppressing invalid association cycles between homologous labeled pathological patches and maximizing the conditional association probability of unlabeled patches belonging to a certain class,respectively.Comparation with existing methods on several performance evaluation metrics demonstrates the advantages of Semi-HIC in semi-supervised pathology image classification.(2)To further enhance the ability of neural networks to represent the features of unlabeled pathology images,based on the previous work,a novel semi-supervised method for pathology image classification based on local augmentation consistency called Semi-LAC is further proposed.Specifically,this method first designs a novel local augmentation technique by randomly sampling local pathological patches from pathology images,and then performs image geometric transformation on these patches,thus ensuring the diversity of augmented pathology images and avoiding the degradation of image classification accuracy due to introduce unimportant regions in pathology images.Furthermore,the directional prediction consistency loss and directional local feature consistency loss are proposed to constrain the consistency relationship between the original pathology images and the augmented images at classification prediction results and local feature-cubes levels,respectively,which also make the prediction result or local feature with lower confidence to aligns the one with high confidence.In this way,the neural network can learn more expressive features from the pathology images and effectively improves the accuracy of semi-supervised classification of pathology images.The experimental results show that the Semi-LAC method can significantly improve the semi-supervised classification performance of pathology images compared with existing methods including Semi-HIC.(3)In addition to the aforementioned studies,a novel semi-supervised method called Semi-NuD is also proposed in this study for nuclei detection.This method introduces a globally consistency regularization loss function for driving the predicted results of the original unlabeled images to be consistent with the perturbed images,thus making contextual information at the pathology image level is fully used.Subsequently,a local adversarial learning technique for pathological patches is also proposed to further enhances the local spatial consistency of the nuclei detection results by using the local nuclei location probability maps generated by detection network and the respective ground-truth for adversarial training.Experimental results on multiple datasets show that the Semi-NuD method achieves excellent performance in the semisupervised nuclei detection for pathology image.