Analysis And Processing Of Medical Tumor Image Based On Deep Neural Networks

With the development of industrialization and the deterioration of environment,the incidence of malignant tumors is on the rise.There is growing concern of society for the situation.Early detection,diagnosis and treatment of cancer is of great significance for the rehabilitation of patients.At present,the medical diagnosis of cancers mainly relies on biochemical and imaging examinations,therefore computer-aided detection and di-agnosis(CAD)based on medical images has an important impact on the comprehensive diagnosis results of cancer.However,tumor image data has complicated features and regular expressions while labeled data of high quality is relatively scarce.How to effectively obtain the key lesion features from the multi-dimensional tumor image data,and conduct semantic segmentation and target detection on the target image is one of the hot spots and urgent problems in the field of medical image and machine learning.Based on the deep convolutional neural network(CNN)feature extraction mecha-nism and conditional random field(CRF)theory,this dissertation studies the key issues of brain and lung tumor image data including preprocessing,semantic segmentation,target detection.The main works are as follows:1)In order to solve the problems of slow convergence and low segmentation ac-curacy in tumor image segmentation model,an image preprocessing algorithm based on distributed histogram is proposed.The intensity values of tumor image are mapped piecewisely by the histogram-based algorithm to a preset fixed interval,and then well-balancedly redistributed.The algorithm can theoratically guarantee the interpretability of tumor image intensity values after preprocessing.Comparative experimental results show that this algorithm can improve the accuracy and convergence rate of tumor image segmentation model.2)Aiming at the problem of lacking global context information in traditional CNN model,a hybrid network architecture is designed,which can effectively fuse local and global context information.This architecture captures the local complementary infor-mation of two-dimensional tumor images through maximum pooling and average pool-ing layers,and consequently extracts the global context information of images by using full-connected CRF(FCRF).The segmentation performance of the model can be im-proved by fusion of local and global context information.3)For the problem of feature selection bias in three-dimensional tumor image seg-mentation,a three-dimensional deep CNN semantic segmentation model based on focal loss is formulated.The model makes full use of the weight adjustment mechanism of focal loss function to reduce the loss weight for the major categories and improve the loss weight of the minor ones.The function adaptively adjusts the impact of different categories on the final loss and effectively solves the problem of feature selection bias induced by unbalanced distribution of categories.4)An automatic detection model is designed to solve the problem of high false positive rate and high computational complexity in the detection of pulmonary nodules.This model conducts sample amplification by leveraging a three-dimensional genera-tive adversarial network(GAN),which can generate new nodules in samples to bal-ance the distribution discrepancy between nodule and non-nodule categories.It also adopts the Single Shot MultiBoxDetector(SSD)network based on feature reuse to de-tect pulmonary nodules.Compared with existing models,empirical results indicate that the model can effectively improve the detection accuracy and sensitivity,and simplify two-step implementation of traditional nodule detection models to a single step.To sum up,four key issues in the preprocessing,semantic segmentation and detec-tion of medical tumor images are studied in depth in this dissertation,covering equal-ization preprocessing of tumor images,robust feature extraction based on deep CNN,semantic segmentation of tumor regions based on different dimensions and accurate tar-get detection.These studies have played a positive role in promoting the development of semantic segmentation and detection of medical tumor images.