Study On Automatic Segmentation Of Thoracic And Abdominal Lesions Based On Convolutional Neural Network And Its Application In Disease Prediction

Thoracic and Abdominal cancers including Thymic Epithelial Tumors(TETs)and pancreas tumors are seriously threatening life and health of people,which suffer from pessimistic diagnosis and treatment.Based on Computerized Tomography(CT)images,accurate segmenting and analying lesions,and predicting postoperative complications have great clinical significance in assisting doctors to make correct decisions.However,the manual interpretations greatly depend on the visual features of the lesions,affecting by the experience of doctors and lacking of consistency.Thus,it is of great clinical significance to develop computer aided diagnosis algorithms.Recently,deep learning and radiomic methods have greatly developed in the fields of medical image processing.However,due to the various shapes and complex textures of TETs,and the obscure specific textures and structures of pancreas,precise segmenting TETs and pancreas is difficult and challenging.The preoperative prediction of Postoperative Pancreatic Fistula(POPF)by analyzing the medical images has great clinical significance.However,the current clinical prediction methods are subjective and hysteretic.For postoperative complications prediction,radiomics can utilize multi-dimension information,but are limited by manual feature extracting process;deep learning methods can model the heterogeneous features automatically,but are weak in utilizing clinical information.Therefore,base on the automatical segmentation of lesions,combining the advantages of radiomics and deep learning algorithm has potentiality in efficient,accurate and consistent prediction of postoperative complications disease.This dissertation aims to tackle with the challenges in TETs and pancreas segmentation based on CT images,and the preoperative prediction of postoperative complications.The main contributions of the dissertation can be summarized as follows:Facing the difficulties in segmenting TETs,the TETs segmentation algorithm including three-channel pseudo-color preprocessing method and Dense Connection Network(DSC-Net)is proposed: 1.three-channel pseudo-color preprocessing method converts grayscale images to three-channel pseudo-color images by different CT windows,highlight and fuse the characteristics of TETs and neighbouring tissues,which can describe the different tissues and highlight the TETs in more detail;2.The proposed DSC-Net are equipped with dense connection structures in the encoder and skip connections,aiming to reuse multi-level information,and integrate low-level detail and deep-level semantic information to increase the representation ability of TETs.Experiments via enhanced CT image dataset indicate that the proposed TETs segmentation algorithm achieves the superior performance over existing models,and is robust in segmenting CT images with different enhanced phases and different TETs subtypes.The local connection characteristic of convolutions and the semantic gap among different dimension features lead to the suboptimal segmentation performance of DSCNet for several complex cases.In order to further improve the segmentation performance of TETs,Multi-level Global-aware Network(MG-Net)is proposed: 1.the cross-attention block is proposed to enhance the intra-class compactness and inter-class difference of features by computing the correlation between pixels in feature map;2.the cross-attention block is further used in skip connections to establish the semantic bridge for the semantic gap between shallow and deep level features;3.adaptive feature fusion module is proposed to calibrate and fuse the decoder features in spatial dimension,providing abundant and discriminative information for TETs segmentation.Experiments indicate that MG-Net significantly improves the segmentation performance of TETs,and achieves the superior performance over existing models including DSC-Net.Different from TETs,the pancreas is the soft glandular organ with various shapes and locations,and have no specific textures and structures in CT images,making it diffcult in segmentation.In order to percisely segment pancreas utilizing 3D features,the 2.5D segmentation algorithm is proposed to explicitly project the 3D information into 2D planes,and guide the 2D segmentation model to learn the variable in the marginal regions of pancreas,enhancing the ability of 2D model in modeling 3D features without high computational requirements.In the same times,the Multi-Attention Dual Context Network(MADC-Net)is designed to effectively fuse multi-level features and locate pancreas.Spatial/channel attention blocks and multi-level feaure fusion module are utilized to effectively model and fuse the discriminative features,which can further characterize details of pancreas combined with 2.5D segmentation algorithm.Experiments with multi-center dataset suggest that 2.5D MADC-Net superiorly performed for pancreas segmentation over exciting models,and is robust on external validation set.In order to preoperatively predict POPF,a preoperative POPF prediction model is proposed based on the previous study in pancreas segmentation,using multi-dimension features extract by deep learning and radiomic algorithms: 1.the high-level deep learning features extracted by transfor learning are combined with low-level radiomic features,clinical and the laboratory information,providing abundant and comprehensive characteristics for POPF prediction;2.multistep feature selection algorithm is desgined to select independent and discriminative features from aforementioned features;3.support vector machine is utilized construct POPF prediction model by comparing it with others classification models.Experiments with multi-center dataset suggest that the POPF prediction model is efficient and stability for all centers,warning surgeons taking timely preventive programs,operation and nursing mode for patients with high risk of POPF,indicating the clinical potential of POPF prediction model.