Study Of Tumor Image Segmentation And Classification Methods Based On Deep Learning

Malignant tumor has become one of the major public health problems that harm the health of residents in China as well as in the world.and its morbidity and mortality exhibit an increasing trend year by year,which presents great challenges to cancer diagnosis and treatment.As medical imaging advances rapidly,automatic processing of medical imaging data of patients by using tumor imaging analysis technology can contribute to accurate and efficient tumor segmentation and classification,which provides important guidance for clinical tumor diagnosis and making treatment plans.Machine learning is a common technology used in early studies of tumor imaging analysis,which presents suboptimal performance in tumor segmentation and classification since artificial features are difficult to extract and cannot explore comprehensive information of tumors.In recent years,deep learning technology represented by convolutional neural network has developed tremendously,and has been successfully applied to tumor imaging analysis.However,due to limited tumor imaging data and complicated tumor shape and structure,studies of tumor segmentation and classification based on deep learning still face huge challenges.To this end,by focusing on gliomas segmentation task and histological subtype classification task of non-small cell lung cancer,the architectural design of neural network and training strategies are actively explored,and thus several novel deep learning methods are proposed to effectively enhance the performance of tumor segmentation and classification.The main contributions of this dissertation are as follows:(1)To address the problem of suboptimal segmentation performance caused by significant difference of tumor shape and size,a deep learning methods GS-UNet is proposed based on improved U-Net,which adopts Up-skip Connection to encode more tumor information of features in lower layers during decoding process.and simultaneously utilizes improved Inception structure to extract more comprehensive features,leading to more accurate tumor segmentation.In addition,we propose an efficient cascaded model training strategy for segmentation subtasks of different tumor regions,which progressively improves the segmentation performance for core and enhancing tumor by using prior knowledge of complete tumor.In this paper,the superiority of GS-UNet in glioma segmentation task is verified by comparing with existing segmentation methods on multiple public datasets.(2)In order to solve the difficulty of extracting discriminative tumor features with limited CT images,we propose a histological subtype classification method RAFENet for non-small cell lung cancer based on reconstruction-assisted feature encoding network,which adopts auxiliary reconstruction task of axial CT images to explore shared information between different tasks,and thus remarkably enhances the generalization ability of the model.At the same time,RAFENet designs a task-specific feature encoding module on the basis of shared encoder,which utilizes cascaded crosslievel non-local blocks to gradually extract multi-level features closely related to histological subtype classification task.In addition,we further propose feature consistency and prediction consistency loss functions for image reconstruction,which provides more effective regularization for the training of RAFENet and thus contributes to more accurate histological subtype classification of lung cancer.In this paper,we validate the effectiveness of RAFENet using multiple evaluation metrics.and the results shows that RAFENet can effectively extract more discriminative tumor features and thus dramatically improve the histological subtype classification performance of non-small cell lung cancer.(3)As CT images in axial view cannot capture 3D spatial information of tumor.on the basis of previous research,we further propose a novel reconstruction-assisted multi-view learning methods MV-RAFENet,which jointly uses CT images of axial view.coronal view and sagittal view to sufficiently leverage characteristics of tumor shape and structure in different views,and meanwhile adopts auxiliary reconstruction task and corresponding multi-view reconstruction loss for improving generalization ability of the model.To efficiently explore the consistent and complementary information among various views,MV-RAFENet utilizes feature decomposition module based on attention mechanism to decompose features from different views to view-common features and view-specific features,and then leverages view consistency loss function to encourage view-common features to be consistent.After that,we utilize adaptive feature fusion module to fuse features of different views with the aim of effectively integrating multi-view information.Experimental results show that MVRAFENet achieves superior histological subtype classification performance than existing methods on several non-small cell cancer datasets.