BtrflyXt:基于MLP与CNN混合结构的脊柱椎体定位网络  

BtrflyXt: A Vertebral Body Localization Network Based on a Hybrid Structure of MLP and CNN

作　　者：张梦斯 李金迪 周迪斌[1] 张陶界 ZHANG Mengsi;LI Jindi;ZHOU Dibin;ZHANG Taojie(School of Information Science and Technology,Hangzhou Normal University,Hangzhou 311121,China)

机构地区：[1]杭州师范大学信息科学与技术学院,浙江杭州311121

出　　处：《杭州师范大学学报（自然科学版）》2023年第6期581-589,共9页Journal of Hangzhou Normal University(Natural Science Edition)

基　　金：浙江省自然科学基金联合重点项目(U21A20466)。

摘　　要：人体脊柱的椎体自动化定位是辅助医生进行病理分析、诊断及评估的重要步骤.但因个体差异、脊柱解剖结构复杂、CT呈现视野范围不同等原因,椎体自动定位困难较大.现有基于深度学习的椎体定位模型大多需要巨大的计算开销,无法应用于临床医学任务.文章提出一种卷积神经网络(convolutional neural networks, CNN)与多层感知机(multilayer perceptron, MLP)混合的BtrflyXt网络结构,使用分组卷积与通道压缩降低模型的参数量,同时在MLP阶段引入注意力机制,搭建了一种端到端、轻量级的椎体中心点定位(与标号)网络,实现了在压缩模型88%参数的条件下椎体识别率提高8%的任务.The automatic localization of vertebral bodies in the human spine is a crucial step in assisting doctors with pathological analysis,diagnosis,and assessment.However,due to individual variations,the complex anatomy of the spine,and variations in the field of view in CT scans,automatic vertebral body localization presents significant challenges.Many existing deep learning-based vertebral localization models require enormous computational resources and are not suitable for clinical medical tasks.This study introduces a novel network architecture called BtrflyXt,which combines convolutional neural networks(CNN)with multilayer perceptron(MLP).This structure reduces the model's parameter count using grouped convolutions and channel compression.Additionally,attention mechanisms are introduced during MLP stage to construct an end-to-end,lightweight network for vertebral center point localization(and labeling).This network achieves task of 8%improvement in vertebral recognition accuracy while reducing the model's parameters by 88%.

关 键 词：椎体定位 多层感知机 卷积神经网络 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]