Machine Learning Based Medical Image Translation And Its Clinical Application

Medical imaging plays an increasingly important role in clinical diagnosis and treatment that various modalities of medical imaging can assist clinicians in diagnosing disease,implementing treatment,and assessing prognosis in a more comprehensive manner.However,multiple imaging scans may cause economic burden to the patients who require multiple imaging examinations.Some patients may not be suitable for certain imaging techniques necessary for disease diagnosis due to their own health reasons.As there is natural correlation between medical images of different modalities of the same patient,medical imaging translation algorithms can be a possible solution to the above-mentioned problems.When a modality of medical imaging for a patient has already been scanned,medical imaging translation algorithms can be used to synthesize medical images of another specific modality required for clinical treatment based on the existing images,thus avoiding the need to actually perform medical imaging scan for this specific modality.Due to the different imaging principles of different modalities of medical images,it is a great challenge to interconvert medical images of different modalities which is clinically available.Magnetic Resonance Imaging(MRI)and Computed Tomography(CT)are the two most common clinical imaging techniques.In this thesis,we investigated the modality translation between MRI and CT images and its clinical application based on deep learning approaches.This thesis first experimentally compares the accuracy and the quality of the synthesized images in MRI and CT interconversion based on supervised and unsupervised deep learning methods respectively.These two method were trained with the same datasets and training setting.The results show that the pixel value curve tendencies of the images synthesized by supervised deep learning are closer to those of real medical images.The unsupervised learning-based method synthesizes rich but possibly erroneous image contrast information.The quantitative measurements of the synthesized images by the supervised learning model are better than those of the unsupervised learning model.In the thesis,we proposed a deep convolutional neural network based on residual learning to generate synthetic CT images from MRI images for the clinical application of MRI-based radiotherapy.CT images is related with electron density(ED)information which is critical for radiation dose calculation in radiotherapy.Since MRI images lacks ED information,generating synthetic CT could make it possible for MRIbased radiation planning.Multiple deep residual learning units are introduced to the imaging translation model,which makes the training process easier to converge and can take full use of the features in input images.The CT images generated based on this model can be used for radiotherapy planning.Clinical diagnosis based on medical images translation can reduce the clinical reliance on the actual scanning of multiple imaging techniques.In this paper,we explored the application of MRI image based CT image synthesis for clinical diagnosis of spine diseases.In the experiment,we synthesized MRI spine images into corresponding CT images by a proposed U-shaped deep learning network,and evaluate the quality of the synthesized images from both algorithmic and clinical perspectives.Clinical radiologists performed diagnosis based on real CT and synthetic CT images,respectively,and the results showed that the CT images synthesized based on the proposed network have the potential to diagnose osteoproliferation?This experiment provides a referable technical route for clinical diagnosis based on medical image modality conversion.