Research On Colorectal Polyp Image Segmentation Algorithm Based On Deep Learning

Deep learning technology is widely used in the field of medical image segmentation,and has certain advantages in accuracy and real-time performance compared with traditional medical image segmentation algorithms.Among them,colorectal polyp segmentation technology plays a key role in the early treatment of colorectal cancer,which can effectively reduce the incidence of colorectal cancer.Polyp segmentation is challenging due to the variable size,color,and texture of colorectal polyps,and the unclear boundary between polyps and surrounding mucosa.In order to solve the problems existing in the actual segmentation scene,this paper proposes two improved algorithms based on Deeplabv3+ and Double UNet respectively.The main work is as follows:(1)A data preprocessing method for different datasets and a post-processing method for segmentation results are proposed.The data preprocessing method includes the colorectal polyp image de-reflection method and the data augmentation method of all datasets.The colorectal polyp image de-reflection method reduces the impact of the reflective area in the colorectal polyp image on the training process,while the data augmentation method solves the problem.To solve the problem of the small number of samples in medical image datasets,a single image can be expanded to 26 different images through different image processing methods;the post-processing methods of segmentation results include two methods: conditional random field(CRF)and test time augmentation(TTA).It is proved that the post-processing method proposed in this paper can refine the final segmentation result and improve the segmentation accuracy.(2)A model-lightweight colorectal polyp segmentation algorithm is proposed.The algorithm uses the Deeplabv3+ network as the benchmark model,and uses the lightweight backbone model Mobile Netv2 for feature extraction,which greatly reduces the number of parameters of the model and improves the training speed of the model.In addition,a BAM attention module is introduced in the encoder to solve the problem that some pixels in colorectal polyp images are difficult to accurately segment.When conducting experiments on Kvasir-SEG dataset and CVC-Clinic DB dataset,compared with the benchmark network Deeplabv3+,the algorithm proposed in this paper has higher segmentation accuracy,fewer parameters,and shorter training time.(3)An improved DoubleUNet network segmentation algorithm is proposed,which first introduces an attention mechanism in the decoder part of the Double UNet network,and replaces the Atrous Spatial Pooling Pyramid(ASPP)module in the network with a densely connected Atrous Spatial Pooling Pyramid(Dense ASPP)module is used to improve the ability of the network to extract features.Finally,in order to improve the segmentation accuracy of small objects,the Focal Tversky Loss function is proposed as the loss function of this algorithm.The accuracy rates of the algorithm in the Kvasir SEG,CVC-Clinic DB,ETIS-Larib,ISIC,DSB dataset tests are 0.9530,0.9642,0.8157,0.9503,and 0.9641,respectively,while the accuracy rates of the Double UNet algorithm in the above datasets are 0.9394,0.9592,0.8007,0.9459,0.9496.The experimental results show that the algorithm in this paper has a better segmentation effect than the Double UNet algorithm,which can effectively assist physicians to remove abnormal colorectal tissue,thereby reducing the probability of polyp canceration,and can be applied to other medical image segmentation tasks.