Research On Medical Image Segmentation Based On Deep Learning

The main purpose of modern medical imaging technology is to improve the accuracy and reliability of clinical diagnosis.With the development and progress of science and technology,the relationship between medical imaging technology and artificial intelligence is becoming closer and closer,and computer-aided diagnostic systems are helping doctors to analyze pathogenesis more intelligently.Due to the variety of medical image types,medical image segmentation tasks based on deep neural networks have become a challenging task.This paper deeply studies and proposes medical image segmentation networks based on deep learning for different segmentation tasks,aiming to improve the segmentation accuracy of the network model.The specific research contents are as follows:(1)A skin lesion segmentation network with adaptive channel-context-aware pyramid attention and global feature fusionThis paper proposes a new adaptive channel-context-aware pyramid fusion network(ACCPG-Net)based on a U-shaped structure for accurate segmentation of skin lesions.Unlike other networks,the proposed network focuses more on the adaptability of feature extraction and the globality of feature fusion.Specifically,a lightweight attention extraction module and global fusion module were added to the encoder and decoder to improve segmentation accuracy.First,the adaptive channel-context-aware pyramid attention(ACCAPA)module was embedded in the encoder.This module models the feature regions of skin lesions in a multidimensional way by dynamically adapting to channel information,context information,and global structural information.Then,the global feature fusion(GFF)module was embedded in the decoder to enhance the semantic information interaction of feature maps at different levels between the encoder and decoder.Comparative experiments and ablation experiments on four public skin disease datasets show that ACCPG-Net has better segmentation performance and generalization ability than other popular methods.(2)A Dual-branch multi-information aggregation network with transformer and convolution for polyp segmentationAbnormal mucosal polyps distributed within the large intestine can easily transform into colon and rectal cancer.Therefore,early detection and screening of polyps are of great significance for the clinical diagnosis of intestinal cancer.However,the location distribution of polyps in the intestine is complex,and intestinal images often contain a lot of noise and reflection,which makes it difficult to distinguish polyp features from normal intestinal tissue features and leads to inter-class interference.To solve these problems,this paper proposes a parallel branch multi-information aggregation network(DBMIA-Net)based on Transformer and CNN for efficient and accurate segmentation of polyps.The paper designs an adaptive channel graph convolution(ACGC)feature extraction module,a global information(GIA),and an edge information(EIA)feature aggregation module.The main function of ACGC is to extract deep features in the encoder and fully represent the potential correlation information between channels.GIA and EIA respectively fuse the feature maps of the dual main network from the perspective of global features and edge features,to better locate the distribution of polyps.The experimental results on five publicly available intestinal polyp datasets show that DBMIA-Net has excellent segmentation performance and generalization ability for polyps.(3)A semantic selection aggregation network for COVID-19 lung infection segmentationSegmentation of infected areas from chest CT images of COVID-19 patients has significant clinical importance for diagnosis and quantitative analysis,as the accuracy of image diagnosis is much higher than that of reverse transcription polymerase chain reaction(RT-PCR)detection.However,the accuracy of existing methods for the segmentation of ground glass opacity(GGO)still needs to be improved.This paper proposes a network for segmenting the lung infection area of COVID-19.First,a dynamic atrous convolution reception(DAR)module is used to extract and fuse high-level semantic information at multi-scales in the deep layers.Then,the semantic selection aggregation(SSA)module is used to select and remove redundant and useless information from each layer’s feature information,achieving precise segmentation of the lung infection area.Experimental results show that the proposed method has excellent segmentation performance and outperforms some popular segmentation models.