Supervised And Semi-supervised Methods For Medical Image Semantic Segmentation

In recent years,deep learning has been widely used in image processing.In clinical diagnosis and surgical planning,it is essential to estimate the targets from medical images accurately.Therefore,many scholars have proposed novel deep learning methods for medical image segmentation.However,these deep learning methods still have some limitations,which are shown as follows:(1)the lesion regions have great randomness and irregular boundaries.(2)The labelled medical images for diseases and organ tissues are scarce.(3)In the capturing processing,the medical images may be unclear due to spatial aliasing,noise and other factors.To solve these problems,thesis constructs a corresponding deep learning framework and conducts ablation and comparison experiments on medical image data sets.Therefore,the main contents of thesis are as follows:1.To solve the segmentation difficulty of tumor regions with randomness and irregularity,a novel residual multi-level and multi-scale semantic segmentation framework is proposed,which stacks three residual multi-scale segmentation networks for brain tumour semantic segmentation,where the first residual multi-scale segmentation network is used to provide position prior information,the second residual multi-scale segmentation network is used to accomplish the coarse-level segmentation,and the third residual multi-scale segmentation network is used to refine the coarse-level segmentation.The residual multi-scale segmentation network embedded the residual multi-scale feature fusion module into the feature extraction part to assist the framework in ensuring the brain tumour regions and obtain better segmentation performance.Furthermore,the residual multi-level and multi-scale semantic segmentation framework introduces a novel boundary consistency loss to capture edge information of brain tumour lesion regions.2.In order to solve the scarcity problem of labeled medical images,a novel semi-supervised semantic segmentation framework is proposed,which comprises global feature extraction branches,and local feature extraction branches,encouraging two branches to generate predictions through collaborative cross-learning and minimizing the differences of predictions through implicit consistent regularization.Meanwhile,the semi-supervised segmentation framework could use local convolution operation and Transformer to capture features comprehensively.Moreover,the framework adopts the dual attention module,which could capture the relative position information of heart structure.Furthermore,the semi-supervised segmentation framework also adopts active contour loss to improve the heart contour recognition ability of the model.3.The thesis conducted extensive experiments to evaluate the performance of residual multi-level and multi-scale semantic segmentation framework and semi-supervised semantic segmentation framework on the Bra TS2015 and ACDC2017 datasets,respectively.Compared with other advanced methods,the residual multi-level and multi-scale semantic segmentation framework and the semi-supervised semantic segmentation framework through collaborative cross-learning have accurate recognition ability.