Research On Organ Segmentation Algorithm In Brain Tumor Region Based On Deep Learning

Brain tumor is a common malignant tumor,which is difficult to treat and extremely lethal due to its special location.Radiation therapy is the most common modality in the clinical treatment of brain tumors.When formulating a radiation therapy plan,it is necessary to accurately outline the target area of brain tumor blocking organs to ensure that the tumor area is applied correct and sufficient radioactive amount and ensure that the health tissue is in a safe range.At the same time,the precise outline of the brain tumor blocking the clinical target volume of brain tumors has important reference value to determine the determination of clinical target volumes of brain tumors（CTV）regions When delineating the results,experienced radiation oncologists manually annotate slice by slice,which is not only cumbersome but also labor-intensive.Therefore,in order to improve the delineation efficiency and reduce the workload of doctors,it is of great research value to develop an accurate and effective automatic segmentation algorithm.However,many state-of-the-art automatic segmentation methods for medical images cannot obtain accurate results due to factors such as large differences in the shape and size of brain tumor barrier structures,blurred structure boundaries,unbalanced front-toback background ratios,and low contrast of different structures.In order to solve the above problems,this paper proposes a High-and-Multi Resolution Network（HMRNet）,which extracts multi-scale feature information while maintaining high-resolution context information so that the network can learn more robust feature information.Furthermore,considering the complementarity between the feature maps in the dual branches of the proposed network,this paper further designs a bidirectional attention calibration module,the feature maps in the dual branches generate a spatial attention weight map to calibrate the complementary branch feature maps,which enhances the model’s attention to the target region.This paper also proposes a false negative（FN）weighted loss function to increase the recall rate of the model to reduce the false negative segmentation results caused by the imbalance between the front and rear backgrounds.Finally,this paper adopts the Coarse-to-Fine two-stage segmentation strategy.The first stage is coarse segmentation with structural localization.In the second stage,a highresolution and multi-resolution dual-branch fusion neural network based on bidirectional calibrated attention module is used.Finely segment each type of structure separately to obtain more accurate segmentation results.The experimental results on the collected brain tumor isolation dataset show that: 1)the Coarse-to-Fine two-stage segmentation strategy proposed in this paper is largely superior to the strategy of segmenting all structures with only one stage;2)the dual-stage segmentation strategy proposed in this paper The branched structure of HMRNet can maintain high-resolution feature information and effectively improve the segmentation accuracy of lamellar structures;3)The performance of the bidirectional calibrated attention module proposed in this paper is better than the attention mechanism that uses its own features for self-calibration.