Research On Medical Image Segmentation Algorithms Based On Boundary Awareness And Feature Complementarity

Medical image segmentation plays a crucial role in smart healthcare.Accurate pixel-level segmentation of organs and lesions from medical images can effectively aid doctors in diagnosing and monitoring diseases.Deep learning,extensively applied to medical image segmentation has achieved remarkably owing to its excellent feature extraction capabilities.Medical image segmentation,however,based on deep learning still faces great challenges due to the presence of targets with blurred boundaries,large scale variations,and cluttered backgrounds.Therefore,the extraction of boundary information and deep features is crucial for enhancing the effectiveness of medical image segmentation.This dissertation conducting research at both theoretical and technical levels approaches medical image segmentation from the perspective of boundary awareness and feature complementarity.1.Polyp segmentation algorithm based on boundary learning and enhancementIn the process of feature extraction,U-shaped convolutional neural networks often encounter the issue of losing detailed information,resulting in inaccurate boundary localization.To improve segmentation accuracy,this dissertation proposes a polyp segmentation algorithm based on Boundary Learning and Enhancement Network(BLE-Net).In encoding stage,the network gradually embeds detailed information from shallow levels into high-level features through boundary learning module,mitigating the loss of details.In decoding stage,a boundary enhancement module is employed to progressively excavate fuzzy boundary clues.By improving boundary awareness during the encoding-decoding stages,BLE-Net achieves accurate boundary localization and boosts segmentation performance.This dissertation presents comparative experiments and performance evaluations on five polyp datasets,with results demonstrating that BLE-Net exhibits outstanding segmentation performance.2.Medical image segmentation algorithm based on region-boundary interactionCorrelations between the target region and the boundary is usually,ignored by existing methods,causing their boundary localization ability limited.In addition,when the target scale changes greatly,these methods have insufficient feature representation,affecting their segmentation effects.Thus,this dissertation proposes a framework based on the interaction of region detection and boundary localization,called CRB-Net.First,CRB-Net applies with two multi-level adaptive feature learning modules to extract discriminative multi-scale features for region detection and boundary localization,respectively.Secondly,a collaborative multi-step refinement module is constructed to mine correlations between the region and boundary.Finally,a parallel iterative decoder is designed to aggregate multi-scale features extracted by the above two modules in the iterative path,improving its boundary localization ability and enhancing segmentation performance of the target region.CRB-Net is evaluated on relevant datasets such as polyps,skin lesion,and nuclei,and the algorithm significantly improves the accuracy of boundary depiction and segmentation performance of multi-scale targets.3.Medical image segmentation algorithm based on complementary and contrastive featuresThe background of medical images has complex characteristics such as low contrast and clutter,which affects segmentation results.In addition,small scale of annotation medical image datasets greatly limits the generalization performance of models.To tackle that,this dissertation proposes a medical image segmentation network based on complementary and contrastive features,called CCNet,and a synthetic data augmentation strategy.First,CCNet constructs a complementary feature extraction module that employs a parallel structure based on Transformer and convolutional neural networks to learn global and local features,thereby enhancing the model’s capability to distinguish complex backgrounds.Second,to enhance the model’s global feature representation,a global context refinement module is employed to incorporate the abstract semantic information of local features into global features.Finally,a mutual attentive module is designed to capture foreground and background contrastive features,enabling the model to overcome interference from complex backgrounds.Moreover,a synthetic dataset based on generative adversarial networks is constructed to expand the scale of training data and increase the diversity of training data,which enhances the generalization capability of the model.Experimental results show that CCNet achieves advanced segmentation performance on datasets from polyps,skin lesions,and nuclei segmentation,demonstrating that the algorithm has strong resistance to interference from complex backgrounds.4.Medical image segmentation algorithm based on global optimization and local enhancementIn terms of the difficulty in segmenting small targets,this dissertation proposes a Locally Enhanced Transformer Network(LET-Net),which combines locally enhanced information with global features to promote a deeper understanding of the image and improve segmentation performance.Firstly,Pyramid Vision Transformer is utilized to extract multi-scale global features.Secondly,a feature-aligned local enhancement module is constructed to align adjacent global features in the spatial dimension and enhance the representation of local information through dense convolution operations.Finally,a progressive local-induced decoder is designed to dynamically guide the decoding process of global features by leveraging local information containing precise details.Moreover,given the imbalance between the number of pixels in small targets and backgrounds,a mutual information loss is devised,which supervises the model in maximizing the preservation of information relevant to small targets while eliminating background noise during training.This dissertation conducts quantitative comparisons and qualitative analyses of LET-Net on polyp and breast lesion datasets.Experimental results show that the algorithm achieves significant performance gains and has excellent capabilities in small target segmentation.In summary,this dissertation tackles the challenges encountered in medical image segmentation,such as inaccurate boundary localization,varied target sizes,and complex backgrounds.Starting from boundary awareness and feature complementarity,this dissertation conducts an active exploration of the feature representation of medical images,model structure,and training optimization,and proposes novel algorithms,effectively enhancing the performance in polyp,nuclei,skin lesion,and breast lesion segmentation.