Deep Learning Models In Medical Image Segmentation

Computer-aided methods are critical for medical image analysis.Recently,deep learning models have attracted intensive research interests and achieved remarkable success.However,the application of deep learning in medical image segmentation still remains challenging.The ongoing issues and challenges are fourfold: firstly,different from natural image segmentation,extremely high segmentation performance is required for medical image segmentation tasks.Secondly,Scanning costs,ethics issues,and the protection of patients’ privacy also pose challenges to the acquisition and publication of medical images.Finally,the distribution of acquired images may differ significantly due to different hospitals,scanner vendors,imaging protocols,patient populations,etc.As the distribution of domain changes,a well-trained system may fail to produce precise predictions for unseen data with the domain shift problem.To tackle these issues,this article focuses on deep learning in medical image segmentation,addresses the high-precision segmentation in medical images,the challenge of data scarcity,and the challenge of domain shift.The main contents of the thesis are as follows:Firstly,this paper studies improving the performance of diagnosis and segmentation.From the perspective of effectively utilizing multi-task knowledge among the labeled data,this paper proposes a cascade knowledge diffusion network that adaptively diffuses the multi-task knowledge.Extensive evaluations and comparisons with state-ofthe-art methods on two skin lesion dataset demonstrate the superior performance of our algorithm.Secondly,this paper studies the problem of data scarcity.From the perspective of effectively using the labeled data and the unlabeled data,a semi-supervised segmentation algorithm with learnable uncertainty is proposed.In summary,we propose a two-stage network with pseudo-mask guided feature aggregation(PG-FANet)and a learnable uncertainty modeling and measurement mechanism.The aggregation mechanisms enable multi-scale and multi-stage feature integration,avoiding the impact of feature incompatibilities in conventional U-shape skip connections.The learnable uncertainty modeling and measurement mechanism penalize both intra-and inter-uncertainties and perturbations in the teacher-student architecture.Comprehensive experimental results over two public histological datasets show that our algorithm outperforms other fully supervised methods,and our semi-supervised learning framework yields competitive performance with a limited amount of labeled data.Furthermore,this paper addresses the problem of domain shift caused by different medical scanner vendors.From the perspective of effectively utilizing the cross-site data,we propose a domain adaptation based self-correction model(DASC-Net).The model,which alleviates the domain shifts,includes an attention and feature domain enhanced domain adaptation model and a dual-domain enhanced self-correction learning algorithm.Extensive experiments using three public COVID-19 CT datasets and comprehensive evaluations demonstrate improved segmentation results over state-of-the-art methods for medical image segmentation and domain adaptation.Finally,to augment the data for segmentation models and solve the problem of data scarcity,this paper proposes a free-form medical image synthesis algorithm from the perspective of generating a large amount of diversified data.In summary,a richer generative adversarial network for free-form medical image synthesis in computed tomography images is proposed,which supports the analytics of various lesions of interests and clinical needs,and improves the authenticity of synthetic images.We evaluate the effectiveness of the proposed generative network by a segmentation task fed with synthetic images.Comprehensive evaluations of the synthesis results are performed on a wide range of public CT image datasets covering the liver,kidney tumors,and lung nodules.The qualitative and quantitative evaluations and ablation studies demonstrate improved synthesizing results over advanced medical image synthesis methods.In summary,this paper takes deep learning application in medical image segmentation as the research goal.This work proposes deep learning models with high segmentation performance,addresses data scarcity and domain shifts,and demonstrates the effectiveness of the algorithms.This work enriches deep learning research on semi-supervised learning,domain adaptation models,attention mechanisms,and generative adversarial models.This research bridges the gap between deep learning theory and medical image segmentation applications,provides guidance for deep learning in medical image analysis applications,and has a certain reference value for the practice of computer-assisted clinical applications.