Research On Lung CT Image Segmentation Algorithm Based On Deep Learnin

Computed tomography(CT)is consistently employed as the primary imaging examination and preoperative diagnosis approach for lung diseases such as lung cancer and COVID-19,which gives doctors great help and effectively promotes the development of medical imaging research.However,when utilizing deep learning methods to segment CT images,there are common problems,such as blurred edges,difficult data labeling,heavy reliance on doctors’ subjective experience and complex lesion types.Thus,automatic segmentation of lesion regions from lung CT images is an extremely challenging task.Implementing automatic segmentation of lung CT image lesion regions can help reduce the diagnostic burden of radiologists and effectively improve the efficiency and accuracy of clinical diagnosis.In order to solve the above problems,this thesis takes the encoder-decoder structure as the basic architecture,and carries out the deep learning model research of lung CT image segmentation from the two aspects: improving the utilization rate of unlabeled images and strengthening the network segmentation capability,so as to achieve high-precision segmentation of lung lesion regions.The main research contents completed are as follows:(1)Semi-supervised Dual-task Balanced Fusion Network(DBF-Net): The fuzzy edges of lung lesion regions cause low contrast with normal background regions and make it difficult to obtain labeled data.Thus,a novel semi-supervised dual-task balanced fusion network is proposed.To better extract the features of lung lesion regions,the model utilizes a lightweight double convolution module and a fusiform equilibrium fusion pyramid module to perform the down-sampling operation,while combining a semi-supervised learning strategy to generate reliable pseudo-labels.In addition,this thesis employs an image enhancement method specifically for medical image processing to highlight the location of lesions,so as to prompt the model to extract more visual features and obtain richer pixel information.A series of experimental results show that the proposed model outperforms other contrast segmentation models in terms of sensitivity,specificity and accuracy,and has strong competitiveness.(2)Progressive Dense Residual Fusion Network(PDRF-Net): Fast and accurate segmentation of lesion sites from patients’ lung CT images is of great significance for early diagnosis and monitoring of patients.To this end,a progressive dense fusion network model is proposed.The model introduces dense skip connections to capture multi-level contextual information and compensate for the feature loss problem in network transmission.Efficient aggregated residual modules are designed in the encoding-decoding structure,which combine the vision transformer with the residual block,so that the model can extract richer and finer lesion features from CT images.In addition,this thesis introduces bilateral channel pixel weighted modules to gradually fuse the feature maps output by multiple decoding branches to obtain high-quality prediction results of segmented images.The experimental results show that the proposed PDRF-Net model has better segmentation performance than other mainstream algorithms,and can achieve accurate segmentation of lung lesion regions.(3)Multiscale Feature-fusion Network(MFNet): When performing lung CT image lesion segmentation,there are problems of under-segmentation and unbalanced distribution of small target regions.Based on the basic architecture of U-Net,this thesis proposes a multiscale feature-fusion network model.In order to better focus on lung lesion features,the model is designed with a self-attention pyramid module to unfold the convolutional processing of features at different levels.At the same time,the selfattention mechanism is used to calculate the pixel correlation and enhance the feature saliency.In addition,the up-sampling of the original U-Net is replaced by the upsampling residual module,which reduces the information loss of low-level features based on the residual structure and channel feature excitation,prompting the decoding branch to recover image information layer by layer.While,the information weighted fusion module is introduced at the end of the decoding branch.According to the adaptive adjustment capability,the feature weighted fusion effect is achieved and the misclassification of image background pixels is suppressed.Specific experimental results show that the quantitative evaluation metrics of this model is higher than other comparison models,and it has better lung lesion segmentation performance.