Intelligent Analysis And Application On Magnetic Resonance Imaging Of Pediatric Tumors

Pediatric tumors are the most common solid tumors and an important cause of death in children.Posterior fossa tumors have an overall majority in children with brain tumor aged 4-10 years.Posterior fossa tumors mainly include medulloblastoma,ependymoma,and astrocytoma.These three kinds of tumors have varying degrees of malignancy,and the treatment and prognosis methods are very different.Therefore,early diagnosis and reasonable treatment are particularly important.At present,medical imaging is required for the diagnosis and prognosis treatment of pediatric tumors.Non-radiative magnetic resonance imaging(MRI)technology is more and more widely used in the application of pediatric diseases,and MRI has the advantages of high soft tissue resolution and multi-planar multi-sequence imaging.As a functional magnetic resonance image,diffusion-weighted imaging(DWI)can provide tremendous tissue information and has important guiding significance in tumor diagnosis.The apparent diffusion coefficient(ADC)parameter map of DWI is used in this thesis,which is to analyze the three kinds of posterior fossa tumors,and to segment and classify the tumors.As to provide solutions for the diagnosis and treatment of children's brain tumors,the work of this thesis mainly includes the following three aspects:(1)Pediatric tumor segmentation task.The segmentation task provides the location of the tumor area for subsequent classification task.In this task,the OSTU threshold method is applied to segment the brain parenchyma firstly,and then the fuzzy C-means clustering and level set method are used to get the brain tumors in the brain parenchyma.After segmentation of 386 tumor sections,an precision of 88.5%is achieved.The square patches are generated from segmentation results for the following classification task.(2)Classification of three types of pediatric tumors based on traditional machine learning methods.Firstly,the features are extracted from the tumor area,including histogram and texture features.Feature selection is then used to reduce feature redundancy.Classification task is performed using logistic regression and random forest classifiers.The random forest classifier has better performance,and the overall accuracy rate of classifier is 92.3%.(3)Classification task of pediatric tumors based on deep learning methods.Using transfer learning,the deep learning method is applied to medical images and three strategies are implemented.In addition,the classification performance is improved by data balance method.In order to understand how neural networks work,convolutional neural network which has the best performance is visualized.Finally,various classification models are integrated to obtain a 97%overall classification accuracy.