| Pathological diagnosis is the "golden standard" for judging malignant tumors.Cell(nuclei)detection and segmentation is usually considered as the basic and key precondition step of computer-aided pathological image analysis.However,due to the differences in staining,complex background and human noise interference of pathological images,at the same time,the size,shape and texture of pathological images vary,and the degree of cell adhesion or overlap varies slightly,resulting in accurate and robust detection and segmentation of cell(nuclei)are very challenging.However,the manual feature extraction methods not only require professional knowledge,but also the cost of feature extraction is high,especially the high quality features.Deep convolution network is an important branch of deep learning.It extracts data features layer by layer and combines and transforms them to form higher level semantic features,which has strong modeling ability.In recent years,it has made some achievements and breakthroughs in pathological image analysis.The main research contents of this dissertation are the detection and segmentation of pathological image cell(nuclei)based on convolution neural network and the fine-grained classification and grading algorithm of pathological image slices.The main innovations are as follows:1)A nuclei detection algorithm based on multi-scale convolution network for pathological images is proposed.Due to the variability of cell type and staining,cell adhesion or partial overlap,robust cell(nuclei)detection is usually a difficult problem.To address this issue,a novel multi-scale fully convolutional neural network approach for regression of a density map to robustly detect the nuclei of pathology and microscopy images is proposed.The procedure can be divided into three main stages.Initially,instead of working on the simple dot label space,regression on the proposed structured proximity space for patches is performed so that centers of image patches are explicitly forced to produce larger values than their adjacent areas.Then,several multi-scale fully convolutional regression networks are developed for this task;this will enlarge the receptive field and not only can detect the single,small size cells,but also benefit to detecting cells with big size and overlapping states.In this stage,we copy the full feature maps from the contracting path and merge with the feature maps of the expansive path.This operation will make full use of shallow and deep semantic information of the networks.Finally,morphological operations and maximum/minimum filters are employed and some prior information is introduced to find the centers of the cells more robustly.Our method achieves about 99.25%detection precision and the F1-score is 0.9924 on fluorescence microscopy cell images;about 85.90%detection precision and the F1-score is 0.9020 on lymphocyte cell images and about 78.41%detection precision and the F1-score is 0.8440 on breast histopathological images.Compared with other studies on these thr-ee datasets,the proposed method achieves better detection performance.2)A novel scheme that combined sparse reconstruction and deep convolutional network for segmentation of nuclei in pathological images is proposed.Because of the non-uniformity of the cell image and the complexity of background,it brings great challenges to accurate cell segmentation.Initially,the sparse reconstruction method is employed to roughly remove the background and accentuate the nuclei of pathological images.Then,deep convolutional networks(DCN),cascaded by multi-layer convolution networks,are trained using gradient descent to efficiently segment the nuclei from the background.In this stage,input patches and its corresponding labels are randomly sampled from the pathological images and fed to the training networks.The size of the sampled patches can be flexible,and the proposed method is robust when the times of sampling and the number of feature maps vary in a wide range.Finally,morphological operations and some prior knowledge are introduced to improve the segmentation performance and reduce the errors.Our method achieves about 92.45%pixel-wise segmentation accuracy and the F1-score is 0.8393.This result leads to a promising segmentation performance,equivalent and sometimes surpassing recently published leading alternative segmentation methods with the same benchmark dataset.3)A deep semantic network for nuclei segmentation in pathological images of multiple tissues is proposed.The shape and appearance of nuclei in pathological images of different patients and different tissues differ greatly.In addition,the staining of slices made by different laboratory centers and different staff is different.This leads to a great challenge for accurate segmentation of nuclei in pathological images.The network model proposed in this thesis consists of encoder module,decoder module and atrous convolution module.The encoder module is to obtain the high-level semantic information of the cell image layer by layer and the decoder module that gradually recovers the spatial information.The atrous convolution module is composed of cascade and parallel atrous convolution operations and which can extract multi-scale features and combine them,so that the model gets strong perception ability for small and large nuclei.During the training period,Log-Dice loss and Focal loss are combined.In order to increase the penalty for the smaller prediction result of the Dice coefficient,which is carried out by logarithmic operation and the negative value is taken as the Log-Dice loss.The above optimization is beneficial for the convergence speed.Compared with several recently published semantic segmentation algorithms,the proposed method achieves the best performance on two latest released pathological image datasets.4)A fine-grained classification and grading method of histopathological images based on multi-task deep learning is proposed.The fine-grained classification and grading of pathological images is of great value in clinical application.However,large intra-class variance and small inter-class variance often exist in pathological images.To address this issue,a deep convolution network model based on multi-task end-to-end learning is proposed to achieve accurate fine-grained classification and grading of pathological images.The key steps of the method are as follows.Firstly,online data augmentation and transfer learning strategy are employed to release model overfitting.Secondly,according to the issue that large intra-class variance and small inter-class variance exist in pathological images,multi-class recognition task and verification task of image pairs are combined in the representation learning process;in addition,the prior knowledge(different subclasses with relatively large distance and small distance between the same subclass)is embedded in the process of feature extraction.The experimental results on BreaKHis,invasive breast cancer classification and lymphoma subtype classification datasets show that the proposed method has good fine-grained classification and grading performance.In conclusion,this thesis combines the deep convolution network with sparse representation,morphological operation and other classical methods,take advantage of the merits of the deep convolution network and classical methods for pathological image analysis,and improve the performance of the system.At the same time,this thesis studies the role of encoder and decoder module,atrous convolution module in the semantic segmentation of pathological images and the contribution of the combination of multiple loss functions to the enhancement of the segmentation performance of pathological images.In addition,based on the multi-task learning method,the prior knowledge(different subclasses with relatively large distance and small distance between the same subclass)is effectively embedded in the feature extraction process to improve the fine-grained recognition performance of pathological images. |
PDF Full Text Request