| With the aging of the population,the incidence of prostate cancer(Pca)is on the rise,and threatening the health of elderly men.Early diagnosis is of great guiding significance for subsequent treatment and prognosis evaluation.MRI has the advantages of non-invasive,non-radiation and high-resolution imaging,and has been recognized as the preferred method of prostate examination.Diffusion-weighted imaging(DWI)is the only imaging technique that can noninvasively reflect the water molecular motion in living tissue,and its apparent diffusion coefficient(ADC)can make quantitative analysis of lesions.Although ADC is a non-invasive biomarker in the differential diagnosis,there is considerable overlap in higher and lower PCa and benign tissues,which may be explained by the attenuation of fitting signal strength in the traditional single exponential model,that is,the diffusion of water molecules conforms to the gaussian distribution.However,the microstructure of biological tissue is complex,and the diffusion movement of water molecules no longer corresponds to the gaussian distribution,and starts to deviate from the fitting curve of the single exponential model.Therefore,it is necessary to use a variety of water molecule diffusion models and conduct model comparison to obtain a better fitting model.Therefore,this study aims to integrate the traditional mono-exponential model(Mono)and other non-Gaussian reconstruction methods,including the stretched exponential model(SEM)and diffusion tensor imaging(DTI)model,diffusion kurtosis imaging(DKI)model and intrinsic incoherent motion(IVIM)model were reconstructed.The Levenberg-Marquardt algorithm is used to calculate the parameters of the reconstruction model,and makes a difference study on the fitting of five DWI reconstruction methods in different regions of prostate tissue.In order to meet the needs of hospitals and researchers using reconstruction methods,reconstruction methods are packed into the software,which is applied to process data automatically to obtain parameters of different models for subsequent research.After completing various reconstruction methods,the parameters of different models are obtained.This paper combines the methods of machine learning to improve the diagnostic performance of prostate cancer.In this paper,MR diffusion-weighted images are collected from 39 patients with prostate cancer and 56 benign patients.A total of 17 parameter maps are obtained using five reconstruction models,and then features of each map is extracted,including histogram and texture features.After obtaining image features,machine learning methods are applied to classification task.The SVM and the random forest classifiers are used to classify benign and malignant prostate lesions.The results show that random forest classifier can achieve an AUC value of 0.98 and a high classification performance.In addition,after sorting the importance of features,the study find that the DKI parameter map is an important indicator of tumor classification. |
