Automatic Differentiation Of Prostate Cancer Based On Magnetic Resonance And Histopathology Images

Magnetic Resonance Imaging(MRI)is widely used in clinical examinations.The accuracy of the diagnosis from multi-parametric MRI(mp-MRI)depends on the experiences of radiologists.Data-driven Artificial Intelligence(AI)can be used in the computer-aided diagnosis(CAD)to help the clinicians diagnose the patients.Prostate cancer(PCa)is one of the highest incident cancers in men.Early diagnosis of clinicalsignificant(CS)is important to the treatment of PCa.In this thesis,we explored the AI algorithms in the diagnosis of CS-PCa by using mp-MRI.Firstly,we used radiomics to differentiate the CS-PCa and non-CS PCa(NCS-PCa)from mp-MRI.First order and texture features were extracted from T2 weighted images(T2WI),diffusion weighted images(DWI)with high b-values,and the apparent diffusion coefficient(ADC)maps,respectively.We built models with features from either a single sequence or multi-sequences and compared their performances on differentiation of CS-PCa and NCS-PCa.PROSTATEx data were used to train the model.The results showed that the model built with first-order and texture features DWI achieved the best performance,with an area under the curve(AUC)of 0.85.Then,deep learning(DL)was used to model the same problem.We compared the influences of different pre-processing,network architectures,and loss functions to the performance of the model.To overcome the problem of limited samples,image patches containing lesion were generated as the input of the model.Different weights were assigned to voxels inside/outside the regions of interest(ROIs).NAS Mobile Net with attention mechanism achieved an AUC of 0.91,exceeding that of the radiomics.Finally,we analyzed the tissue microarray images(TMA)of prostate cancer to estimate the regions with different pixel-level Gleason Grade and the region-level Gleason Score of the cases.The Gleason2019 Challenge data were used to train the models.PSP Net and Nest Net were used to segment TMA images by the Gleason Grade before the final Gleason Score was calculated for the patient.Our experimental results showed that our model achieved satisfying results in both Gleason grading and scoring.