Research On Instance Segmentation Method Based On Multi-Level Refinement Fusion Attention

Instance segmentation refers to the segmentation of objects in an image,where each object in the image is separated for further analysis and research.This segmentation method has broad applications in the fields of computer vision and pathological image processing.According to the data report of the International Agency for Research on Cancer of the World Health Organization,colorectal adenocarcinoma is the third most common cancer in the world and the third leading cause of cancer deaths,following lung cancer and stomach cancer.Colorectal adenoma cell instance segmentation refers to the segmentation of colorectal adenoma cell images,where each cell in the image is separated for further analysis and research.This segmentation method is of great importance in the fields of medicine and biology,as it can not only assist doctors in accurately diagnosing patients with diseases,but also greatly improve the efficiency of pathological research.However,due to the complexity of the scale of colorectal adenoma images,the various morphological changes of cancer cells,and the need for strong manual intervention and low accuracy of current segmentation methods in the field of colorectal adenoma cells,precise segmentation of colorectal adenoma cells has become an extremely challenging problem.This article explores the significance of colon adenocarcinoma instance segmentation and investigates current methods proposed for colon adenocarcinoma instance segmentation using the colon adenocarcinoma dataset.Additionally,two novel methods for adenocarcinoma instance segmentation are proposed.The first method proposes a colon gland cell instance segmentation method based on multi-level refinement and self-attention mechanisms.This method improves upon the Mask R-CNN instance segmentation base network by using a feature extraction network with more granularity levels to capture coarse and fine features,and a feature pyramid network to represent these features effectively.The attention module with skip connections is used to enhance global information contact and model context information,and the output features of the attention module are used as inputs to the refinement structure,resulting in improved instance boundary segmentation.On this basis,a second segmentation method was proposed with a focus on the problem of unclear adhesion between cells and confusion with the background.It has higher segmentation accuracy,robustness,and generalization ability,and performs excellently on the Gla S and CRAG datasets.The second method is a glandular cell instance segmentation method based on a multi-scale feature fusion dual-head attention mechanism,using a top-down approach.The neural network has a new spatial-geometric dual-path attention module that not only focuses on the position of cells but also on their geometric shape.With the help of the new fusion module,it can also perceive higher-resolution features,effectively achieving multi-scale feature fusion.The present study used a glandular cell dataset to compare the two methods with other segmentation methods,verifying the superiority of the proposed method,and conducting ablation experiments to fully validate the effectiveness of the designed module.Visual results show that this study can solve the difficult problems of colon adenoma cell instance segmentation and achieve accurate segmentation.