Research On Machine Learning Methods For Medical Image And Its Application In Precise Diagnosis Of Ovarian Cancer

Ovarian cancer is a common malignant tumor of the female reproductive system,its fatality rate ranks first among gynecological malignancies.Among them,epithelial ovarian cancer(EOC)is the main type.Accurate assessment of EOC status before surgery can help patients implement personalized and precise treatment,thus improving their survival rate.Compared with common imaging techniques such as ultrasound and computed tomography,Magnetic Resonance Imaging(MRI)has distinguished advantages in the EOC staging and evaluation of recurrence as well as metastasis,and is the first choice for imaging-based EOC preoperative diagnostic method.However,EOC image manifestations are extremely complex,clinical diagnosis based on the naked eye is highly subjective and easy to misdiagnose.Machine learning methods can mine deep-level features in images from a large number of samples,build models to improve the accuracy of diagnosis,and effectively reduce misdiagnosis.This paper uses machine learning methods to conduct large-scale multi-center researches on a series of problems regarding the segmentation and diagnosis of EOC.The main works are as follows:(1)Aiming at the problem of heavy workload and high difficulty in the manual delineation of EOC lesions for clinicians,Multiple Side-Output Fully Convolutional Network(MSO-Net)is proposed for EOC automatic segmentation.MSO-Net uses multiple side-output modules to decode the extracted features at different levels,thus generating side-output segmentation results containing different scale information,then those results are merged to generate a more accurate final segmentation result.The proposed model can improve the accuracy and stability of EOC delineation and reduce the workload of clinicians.(2)Aiming at the difficulty in clinical differentiation between type Ⅰ and type ⅡEOC,this paper proposes a multimodal MRI-based radiomics-machine learning model.Eight groups of features from multimodal MRI(T2WI FS,DWI,ADC,and CE-T1WI)data are extracted,then support vector machine,decision tree,and linear model are used to construct multi-modal hybrid models,respectively.In addition,the occlusion test is used to identify the critical region for model differential diagnosis,which can potentially assist the positioning of intraoperative frozen slice sampling.(3)Proposing a novel multiple instance convolutional neural network for the identification of EOC and BEOT.The proposed model can learn from the decision-making manners of clinicians to automatically perceive the importance of different MRI modalities,and achieve multi-modal MRI feature fusion based on their importance.In addition,the model also uses strong prior knowledge of tumor distribution as an important reference for decision-making,when making the prediction of one instance,it refers to the predictions of its neighboring instances,thus realize more accurate EOC and BEOT identification.