Radiomics Analysis Of Multiparametric Prostate Magnetic Resonance Imaging

With the rapid development of China's economy,the incidence rate of prostate cancer(PCa)shows the tendency of fast rising.Multiparametric magnetic resonance imaging(mp MRI)is a reliable imaging method for the assessment of prostate cancer.However,subjective interpretation of magnetic resonance images depends heavily on radiologists' expertise and experience,thereby limits its accuracy.Radiomics analysis is a quantitative analysis method,a common way to improve the accuracy and reproducibility of diagnosis.The rise of deep learning techniques brings new ideas to radiomics analysis methods.In this thesis,some key problems of PCa classification based on radiomics features and deep learning were studied,the main contents are as follows.(1)Since the diagnostic information of prostate cancer is widely distributed in different sequences of MR digital images,multi-model features were extracted from different sequences.The diffusion-weighted imaging(DWI)was used to calculate the apparent diffusion coefficient(ADC)map,as a new sequence together with T2 weighted imaging(T2WI),from which radiomics features were extracted.For dynamic contrast enhanced(DCE)sequence,the pharmacokinetic model was fitted with the average concentration of contrast agent in the region of interest,and the pharmacokinetic parameters were obtained as quantitative features.After multi-model feature extraction,a two-stage feature selection method was used to find a feature subset that were helpful in distinguishing PCa.Finally,PCa was classified by using a linear support vector machine.Experimental results showed that the area under the receiver operating characteristic(AUC)was 0.95±0.01.(2)For a relative small data size of prostate magnetic resonance images,the amount of data was not enough to fully train the convolutional neural network(CNN)and the transfer learning concept was used.,First two transfer learning methods,including pretrained CNN model as feature extractor and fine-tuned CNN models,were experimented with single-model magnetic resonance images,then three different multi-model feature-fusion methods were experimented.Experimental results showed that for classification of PCa and non-cancerous tissues under single-modal magnetic resonance images,the pretrained Res Net model achieved the best result,the area under ROC curve was 0.95±0.01 on T2 WI and 0.82±0.02 on ADC,respectively.Among the three multi-model feature fusion methods,the fusion of deep features on T2 WI and manual features on ADC and DCE sequences was the best.The area under ROC curve was 0.97±0.02?...