The Research On Multi-scale Colonoscopy Polyp Image Segmentation Method

Colorectal cancer is the third most common cancer in the world.And it is a public safety issue that threatens human life and health.A study showed that 95% of colorectal cancers are caused by the deterioration of colorectal adenomatous polyps,so it is important to remove colorectal polyps at an early stage to prevent and treat colorectal cancer.In clinical practice,colonoscopy is an effective technique for colorectal polyp detection,which provides important information such as location,hardness,and shape for doctors’ diagnosis and surgery.Traditional and deep learning methods for colonoscopy polyp segmentation are key to preventing and treating colorectal cancer.This paper aims to use deep learning techniques to accurately segment polyp images in colonoscopies.It is showed that polyp segmentation is different from other medical image segmentation.The large size and shape difference of colonoscopy polyps greatly increases the difficulty of segmentation.Achieving accurate segmentation of edge regions is challenging due to their color and texture similarity with the surrounding normal mucosal tissue.This paper works on polyp segmentation in colonoscopy image then proposes modules and networks to address segmentation issues and difficulties.To address the difficulty of polyp segmentation in colonoscopy,the Multi-Attention and Context Network for Polyp Segmentation(MACNet)is proposed by simulating the process of determining the segmentation region by doctors.First,the polyp position is determined and a rough polyp segmentation prediction map is obtained.Then,the false positive and false negative regions are rectified by comparing the non-confident regions with the confident regions,and finally a fine polyp segmentation map is obtained.Specifically,Position Rectify Module(PRM)is proposed to adjust the position of the polyp in the channel direction,horizontal direction and vertical direction of the feature map,which alleviates the problem of inaccurate polyp localization caused by the loss of position information due to multiple downsampling operations in the encoding process.Balanced Attention Module(BAM)focuses on the influence of the polyp area,normal mucosal tissue area and boundary area on the model training in the feature map.In BAM,the attention scores of these three parts are calculated according to the prediction results of the upper layer,and then the feature maps with differentiated attention are obtained by superimposing in the channel direction for purposeful training.Non-local Information Statistical Attention Module calculates the relationship between any two pixels in the feature map,and supplements the full text information of the feature map through long-distance dependence.Focus Module(FM)uses the context inference in the distraction finding phase to find the false positive and false negative regions in the segmentation results,and then uses pixel-by-pixel addition and subtraction in the distraction removal phase to rectify the feature map of the upper layer,obtaining a more accurate feature map for more accurate segmentation.In order to alleviate the problem of large-scale difference of polyps,distraction finding phase uses different scale convolution kernels to obtain feature maps of different sizes,and uses multi-scale feature maps for context inference to achieve attention to polyps of different sizes.Experiments have shown that MACNet performs better in the task of polyp segmentation in colonoscopy images.Based on the excellent image generation capability and the ability to process high-order features of Generative Adversarial Network(GAN),Multi-Scale Generative Adversarial Network for Polyp Segmentation(MSGAN)is further proposed.MSGAN uses MACNet as a generator to generate colonoscopy polyp segmentation prediction maps,builds an independent discriminator to identify whether the input image is a real colonoscopy polyp segmentation image or a generator-generated colonoscopy polyp segmentation prediction map.And MSGAN continuously improves the generator’s generation ability through adversarial training until the discriminator can no longer distinguish.In order to alleviate the instability caused by training with different domain datasets and transfer more gradient information,MSGAN is designed as a multi-scale generative adversarial network,and the discriminator can identify the authenticity of the colonoscopy prediction maps generated by the generator at different scales.Experiments show that MSGAN can better realize the task of polyp segmentation in colonoscopy images.