Research On Multimodal Synthesis And Segmentation Algorithms In Medical Images

In the acquisition process of modern medical images,there are many imaging methods,which often involve many different devices or different parameter settings of the same device.This provides multi-perspective information for expert diagnosis.They can highlight information about different tissues and lesions under the same anatomical structure.However,because the multimodal imaging examination process takes a long time,the images in some modalities may be missing due to uncooperative patients or the acquired images are destroyed by noise or artifacts.On the other hand,the differences in image modalities may cause difficulties for automatic analysis algorithms in medical images,because an image analysis algorithm is often proposed based on a specific image feature,and it is difficult to extend it to other modality images.In order to solve these problems,image synthesis algorithms are widely used in medical image processing to help subsequent medical image analysis algorithms.In this context,thesis proposes deep learning-based synthesis and combined joint segmentation algorithms for the complete sequence synthesis of multimodal MRI and the noise adaptation of optical coherence tomography(OCT)images,respectively.In general,MRI sequences contain multiple image modalities,and most existing image synthesis algorithms cannot flexibly and efficiently process multiple modalities simultaneously.For multi-modal MR image synthesis,thesis proposes a multi-input(and multi-output)structure synthesis method model.It does not rely on complete input modality to synthesize the output,but can use additional inputs to help achieve more realistic images.At the same time,the powerful synthesis ability of adversarial learning is also used to improve the quality of synthesized images.The attention mechanism is introduced into the discriminator to help the discriminator better learn multi-scale features.Gradient detectors better protect tissue structure information in medical images.Experimental results show that the image synthesized by this method has good quality.Among OCT segmentation algorithms,many machine learning-based algorithms assume that the training and test data have the same feature distribution.However,OCT images acquired with different equipment or different parameters of the same equipment have different levels and types of noise.In this thesis,we proposes a noise adaptive synthetic segmentation algorithm.Through adversarial learning,the unlabeled data is transformed into data has the same noise distribution as the labeled data,but with the same content structure.The segmentation results of the unlabeled data can be obtained by using the segmentation network trained by the labeled data to segment the transformed data.The tasks of maintaining the content structure of the data and transforming the noise distribution are completed by two different discriminators with different structures.Experimental results show that this method can achieve good segmentation results.