改进TransUnet模型的脊柱CT图像实例分割方法

打开文本图片集

中图分类号：TP391.4 文献标志码：A 文章编号：1007-2683（2025）06-0001-08

Abstract：A spinal CTimage instance segmentation methodbasedonanimproved TransUnetmodel is proposed toaddress the issueofhighomputationalcomplexityin3Dconvolutionalneuralnetworksinvertebralinstancesegmentationtasks.Themetod achievesinstancesegmentationofachvertebrainthespineonthe2DsagitalplaneofteCTimage.FirstlyasedontheTasUnet network，improvementsaremadebycombiningmulti-scalefeaturefusiontohancethesgmentationacuracyofthemodelforspinal edges.Secondlyamixedatentionfeaturefusionmodule（MAFF）isproposedtorduceduplicatecalculationsinmulti-scalefeature fusionandiprovemodeliciencyFinally，forthetaskofvertebralinstancesegmentation，asagialsliceiterativesegentation methodisproposedtoreduceclasificationerrors invertebralinstancesegmentationresults.Theexperimentalresultsshowthatthe proposed spinal instance segmentation method scored 91. 12% on the dice coefficient，which is 8.47% higher than TransUnet，and the computational cost is reduced by 78.6% compared to the typical 3D convolutional method's Iterative FCN network，proving the effectiveness of the proposed method.

KeyWords：spineCT images；TransUnet model；atention mechanism；iterativesegmentation；image segmentation

0 引言

脊柱不仅支撑和维持着身体器官的结构，还在肌肉骨骼系统的移动中发挥着至关重要的作用。（剩余10890字）