Abdominal Organs And Lesions Segmentation Based On Attention Mechanism And Multi-scale Networks

With the rapid development of artificial intelligence,deep learning is gradually being studied and applied in clinical medicine,and automatic segmentation of medical images is one of its more common applications.Automatic segmentation of CT and MRI images can reduce the workload of doctors,while deep learning can brings faster and more accurate analysis results,and can help doctors diagnose and treat patients’ diseases more effectively.Based on clinical needs and deep learning methods,this paper proposes automatic segmentation methods of prostate and cervical and peripheral organs in response to the actual needs and problems in the radiotherapy of abdominal prostate and cervical cancer.The work of this article is summarized as follows: 1.A multi-scale global convolution 3D MRI prostate segmentation method based on channel attention was proposed to outline the prostate quickly and accurately.Firstly,the classification sub-task is weakened in commonly used segmentation networks,and the global convolution module is introduced into the segmentation network.This can expand the perceptual range of the feature information.At the same time,according to different levels of resolution,global convolution modules with different scales are added.Secondly,considering that different types of low-level features of the encoder module are not necessarily very useful for reconstructing information in the decoder,channel attention is paid to low-level features to correct the contribution of different types of features.Finally,a loss function sensitive to data imbalance is introduced,which can help train more effectively.Analysis of the experimental results shows that this method can obtain better segmentation results.2.A feature fusion 3D MRI cervical cancer segmentation method based on spatial and channel attention is proposed,which can realize automatic segmentation of cervical cancer and assist doctors in cervical cancer radiotherapy.Firstly,in the encoder-decoder segmentation network,the low-level features of different resolutions contain different local and global information,which have a positive effect for the segmentation.Therefore,when performing high-low feature fusion,the outputs of neighboring two-layer encoder blocks are combined as low-level features and fused together.Secondly,in view of the fact that different levels of features may have redundant and unimportant information,the spatial and channel attention are paid to the higher lower-level features,and the more effective channels and spatial features in cross-layer features are emphasized.In order to achieve more effective fusion,in view of the extreme category imbalance of cervical cancer data,weighted dice coefficient loss is used as the training loss function.The above method was used to segment cervical cancer,and good segmentation results were obtained.3.A 3D CT cervix and peripheral multi-organ segmentation method based on hierarchical dilated spatial pyramid convolution was proposed to segment the bladder,rectum,sigmoid colon,and intestine around the cervix to further shrink the region of the cervix,which can relatively reduce the difficulty of segmenting cervical cancer.Firstly,in order to integrate multi-scale global and local feature information,dilated convolution operations with different dilation rate are cascaded to form a dilated spatial pyramid structure.At the same time,considering the inconsistent feature resolution of the encoder,different levels of features use different numbers of dilated convolution,which can avoid the introduction of redundant operations.Secondly,due to the poor separability of edges,especially between bladder and intestine,edge refinement structures are added,and edge feature information in both within-slice and between-slice is explored.Finally,according to the data analysis,the number of pixels between this data category is quite different,so weighted dice coefficient loss is introduced to alleviate the problem of data imbalance.The experimental results show the effectiveness of the above methods,especially in rectum and intestine.