Research On Refinement Analysis Algorithm For Pathological Images

Pathological section is the gold standard for clinical disease diagnosis.Pathologists perform pathological examination and prognosis evaluation by microscopic examination of pathological sections.The whole process is usually time-consuming and challenging for pathologists.In recent years,in the context of the increasing popularity of digital pathological sections,machine learning method has entered the field of pathological,and driving the pathological analysis gradually shift from qualitative analysis to quantitative analysis.Computer-aided digital pathology analysis can help pathologists to overcome the fact that manual diagnosis is susceptible to cognitive ability,subjective experience,fatigue and other factors,and can effectively improve the accuracy and stability of the pathological diagnosis,reduce misdiagnosis and missed diagnosis,which is of great significance for the diagnosis and the choice of treatment options.At present,both traditional machine learning and recently developed deep learning have shown great potential in pathological image analysis,however,because the clinical application of computer-aided pathological diagnosis needs to meet various more refined requirements and the lack of labelled data and other challenges,most existing research methods still cannot meet the requirements of clinical application.Based on this situation,this paper further explores the refinement analysis of pathological images.The main research work are as follows:(1)Aiming at the high efficiency requirement of clinical pathological diagnosis,we propose a novel distributed parallel approach,which adopts both data and model parallel for fast skeletal muscle image segmentation.With a master-worker parallelism manner,the image data in the master is distributed into multiple workers based on the Spark cloud computing platform.On each worker node,we first detect cell contours using a structured random forest contour detector with fast parallel prediction.Then we generate region candidates using superpixel method.Finally,we propose a novel hierarchical tree based region selection algorithm for cell segmentation based on the conditional random field algorithm,which is further parallelized by using multi-core programming.Experiments show that the parallel method proposed in this paper achieves a speed increase of 10 times compared with the stand-alone mode in large-scale skeletal muscle image segmentation.(2)Aiming at the high accuracy requirements of clinical pathological diagnosis,we propose a fine segmentation algorithm for skeletal muscle image based on deep hierarchically-connected networks.The proposed network achieves multi-scale prediction by adding a decoder with independent loss function at different layers of the encoder module,and combines multi-scale prediction results to generate more robust fine segmentation,which effectively solves the problem that the output of the existing end-to-end convolutional neural networks is relatively coarse in cell segmentation.We also propose a two-stage transfer learning strategy to effectively train such deep network.Experiments on the skeletal muscle image dataset demonstrate that the proposed method has significant improvements in timeliness and accuracy compared to other existing methods.(3)Aiming at the lack of a large number of labeled pathological images,we propose a novel histopathology images classification algorithm based on semi-supervised deep linear discriminant analysis.Specifically,we first replace the loss function of the deep neural network with the objective function of linear discriminant analysis to produce features minimizing the intra-class distance yet maximizing the inter-class distance,in order to construct a robust and effective graph Laplacian.Afterwards,we design a new objective function via employing the graph constructed by features of labeled and unlabeled images,and then adopt this objective as the loss function of the deep neural network.Finally,the features generated by the network are used to complete the classification.Validation experiments on the images of skeletal muscle and lung cancer proved that this method was superior to most existing methods.(4)Aiming at the high practicality requirement of clinical pathological diagnosis,we propose a deep learning-based lung cancer survival analysis model.Firstly,an end-to-end cellular feature learning module is constructed by using a deep neural network with global average pooling,and the cellular feathers are aggregated into patient-level feature vectors by using bag of words and locality-constrained linear encoding algorithm.Then the Cox proportional hazards model with an elastic net penalty is proposed for robust feature selection and survival analysis.Finally,A biomarker interpretation module which can help localize the image regions that contribute to the survival model's decision.A large number of verification experiments have proved that the proposed survival analysis model has good predictive ability for the TCGA lung cancer dataset.