一种基于改进的AFFormer网络的脊柱椎体分割方法  

A spinal vertebral segmentation method based on an improved AFFormer network

作　　者：王玉婷 张新峰[1] 郭伟 李相生 刘晓民[1] WANG Yuting;ZHANG Xinfeng;GUO Wei;LI Xiangsheng;LIU Xiaomin(School of Information Science and Technology,Beijing University of Technology,Beijing 100124;Air Force Medical Center,PLA,Beijing 100142)

机构地区：[1]北京工业大学信息科学技术学院,北京100124 [2]中国人民解放军空军特色医学中心,北京100142

出　　处：《北京生物医学工程》2024年第6期566-574,共9页Beijing Biomedical Engineering

基　　金：北京市自然科学基金-海淀原始创新联合基金(重点研究专题)(L222018)资助。

摘　　要：目的 提出一种基于改进的AFFormer模型实现脊柱椎骨的准确分割,辅助医师快速诊断脊柱侧弯情况。方法 本文数据集为全脊柱正位X线影像,图像尺寸约为5 000×8 000,针对图像尺寸大、前景区域小背景区域大、受试者个体差异大但相同部位差异小的数据特点,使用轻量级语义分割模型AFFormer进行脊柱分割。针对深层特征图中损失大量细节信息的现象,在用像素语义对特征中的局部细节进行建模时,在原有的16维度特征图的基础上拼接一个支路输出8通道的特征图实现多尺度特征融合,从而学习到更多的细节信息。数据集为某医院临床的正位全脊柱X线影像146张,使用labelme工具对图片进行像素级注释后,按照8∶1∶1随机划分为训练集(117张)、验证集(15张)和测试集(14张)。训练网络时,使用交叉熵、Dice系数以及增加先验知识的自定义得分函数的加权和作为损失函数,来优化模型训练。在验证集上,使用平均交并比和平均准确度进行检验,进一步调整模型的超参数并初步评估模型从而选择表现最优的模型。结果 使用所提出的方法训练得到的模型在测试集上进行测试,取得了最高的mIoU值(0.867 8)、mAcc值(0.923 2)。结论 本文提出的方法经实验证明其分割性能优于现有的主流分割模型,能够实现脊柱椎骨的精确分割,为辅助脊柱医学诊断提供坚实基础。Objective To propose an improved AFFormer model for accurate segmentation of spinal vertebrae,which assists physicians in quickly diagnosing scoliosis.Methods The dataset in this article is a full spine orthopedic X-ray,with an image size of approximately 5 000×8 000.Considering the characteristics of large image size,small foreground area,large background area,and large individual differences among subjects but small differences in the same part,the light-weight semantic segmentation model AFFormer is used for spinal segmentation.In response to the phenomenon of losing a large amount of detail information in deep feature maps,when modeling local details in features using pixel semantics,a branch output 8-channel feature map is concatenated on the basis of the original 16 dimensional feature map to achieve multi-scale feature fusion,thereby learning more detail information.The dataset consists of 146 clinical orthopedic full spine X-ray images from a certain hospital.After pixel level annotation using the labelme tool,the images are randomly divided into a training set(117 images),a validation set(15 images),and a testing set(14 images)in an 8∶1∶1 ratio.When training the network,the weighted sum of cross entropy,Dice coefficient,and a custom score function that adds prior knowledge is used as the loss function to optimize model training.On the validation set,we use the average intersection to union ratio and average accuracy for testing,further adjust the hyperparameters of the model,and preliminarily evaluate the models to select the best performing model.Results The model trained using the proposed method is tested on the test set and achieves the highest mIoU value(0.867 8) and mAcc value(0.923 2).Conclusions The method proposed in this article has been experimentally proven to have better segmentation performance than existing mainstream segmentation models,and can achieve precise segmentation of the spinal vertebrae,providing a solid foundation for assisting in spinal medical diagnosis.

关 键 词：脊柱分割 辅助医疗 Transformer结构 多尺度特征 

分 类 号：R318.04[医药卫生—生物医学工程]