Researches On Medical Image Segmentation Based On Deep Learning

With the continuous development of diagnostic imaging technology in the medical field,various imaging methods have emerged endlessly.X-ray,tomography(CT)and magnetic resonance imaging(MRI)technologies have been widely used in medical diagnosis.However,relying on large-scale human eyes to observe the diseased area and the potential diseased area will consume a lot of manpower and material resources.How to efficiently detect and segment the diseased tissue through some automated means has become the key.Some traditional machine learning methods are used in medical image segmentation,However,the traditional segmentation method is not very effective due to factors such as partial volume effect,gray unevenness,and insignificant gray difference between different soft tissues.In recent years,the deep learning methods have flourished,and it has also been used in the segmentation of medical images.From the initial convolutional neural network(CNN)to the subsequent full convolutional neural network(FCN),to The emergence of U-Net segmentation models,a large number of segmentation models and algorithms have emerged,and achieved good results.In order to obtain better segmentation accuracy,this article combines some of the previous classic modules and network architecture.Three innovative algorithms are proposed for medical segmentation problems,and good results have been achieved in the corresponding segmentation problems.The main work done in this article is as follows:1.A U-Net segmentation model based on attention mechanism is proposed.The symmetrical encoding and decoding structure U-Net can fully combine the low-resolution information that provides the basis for object category identification and the high-resolution information that provides the basis for accurate segmentation and positioning.On top of this,the channel and space domain attention modules are embedded.Under the action of dual attention,the network optimizes the weight of the target features extracted by the codec module.In the case of highlighting the characteristics of the target area the background area is suppressed and the noise is filtered out,making the background and foreground more discriminative.In addition,it has achieved good segmentation results on the LUNA lung dataset and ISBI cell dataset,which highlights the extraction of key features and improves the segmentation accuracy compared to existing algorithms.2.A Dense Net segmentation model based on full convolution is proposed.U-shaped full convolutional coding and decoding structure and skip connection are used to make the network make better use of context information and fuse feature information of different scales.Dense Net is integrated into the encoding and decoding of U-shaped network.Through the dense connection of network models to reuse features,the extracted features are more detailed,the model parameters are reduced,and the training convergence is accelerated.The N4 ITK algorithm is used in the preprocessing,and the loss function of Dice and cross entropy is used.Finally,the Bra TS 2017 brain tumor data set was used for the experiment.The overall tumor area Dice coefficient reached89.73%,and the accuracy of the LUNA lung data set reached 99.18%,which is better than the existing advanced algorithm.The segmentation is more accurate and detailed.3.A Dense Net segmentation model based on context information is proposed.The Dense Net basic network in the previous chapter's codec network was replaced with a smaller Dense Net to reduce the model size,thereby reducing the amount of calculation and speeding up training.Based on the advantages of the Inception-resnet v2 model,a contextual information extraction module is designed and embedded into the codec.This module includes two parts: dense atrous convolution(DAC)and residual multi-core pooling(RMP).It can get more abstract features and retain more spatial information.Finally,it was tested on the brain tumor data set.The effect of the three tumor segmentation regions was better than the algorithm proposed in the previous chapter,and the indicators on the LUNA lung dataset were slightly better than the previous chapter,which proved that The validity of the context information module.