摘要
目的 探讨深度学习模型在腰椎磁共振T2加权成像(T2WI)矢状图像上全自动识别腰椎间盘退变程度的可行性.方法 回顾性抽取2020年8月至2022年6月于安徽医科大学第三附属医院就诊并行腰椎MRI检查的94例患者的腰椎T2WI图像数据,共获得466个椎间盘,由两名放射科医生手动标注腰椎间盘,将数据随机分为训练集(300个)、调优集(72个)和测试集(94个),首先使用U-Net网络训练椎间盘分割模型,模型评价指标包括Dice系数和交并比(IoU)分数;然后利用SpineNet网络训练分类模型进行评价,评价指标包括准确度、敏感度、特异度、F1分数及ROC曲线.结果 测试集中U-Net模型对腰椎间盘分割的平均Dice系数值及IoU分数分别为0.920、0.853;SpineNet分类模型对腰椎间盘退变分类诊断的准确度、特异度、敏感度分别为0.913、0.912、0.916,ROC曲线分析示,该模型区分腰椎间盘退变轻度vs中度、轻度vs重度、中度vs重度的AUC值分别为0.89、0.95、0.90.结论 深度学习网络对腰椎间盘退变程度的全自动分类是可行的.
Abstract
Objective To investigate the feasibility of a deep learning model for the fully automatic classification of disc degeneration based on lumbar structures on sagittal T2WI images.Methods The lumbar T2WI image data of 94 patients who underwent lumbar spine MRI examination in the Third Affiliated Hospital of Anhui Medical University from August 2020 to June 2022 were retrospectively selected,and 466 discs were obtained.The lumbar intervertebral disc were manually annotated by 2 radiologists on sagittal T2WI images.The data were randomly divided into train set(n=300),validation set(n=72),and test set(n=94).Firstly,a U-Net network was used to train the disc segmentation model.The evaluation indexes of the model included Dice coefficient and IoU score.Then,SpineNet network was used to train the classification model,and the evaluation indexes of the model included accuracy,sensitivity,specificity,Fl score,and ROC curves.Results In the test set,the dice coefficient and IoU values of U-Net model for lumbar disc segmentation were 0.920 and 0.853,respectively.The accuracy,specificity and sensitivity value of SpineNet classification models for lumbar disc degeneration were 0.913,0.912 and 0.916,respectively.The ROC curve analysis showed that the AUC values for distinguishing mild to moderate,mild to serious,and moderate to serious lumbar disc degeneration were 0.89,0.95,and 0.90,respectively.Conclusion It is feasible to realize the fully automatic classification of disc degeneration based on deep learning network.
基金项目
安徽省卫生健康委科研项目(AHWJ2021bl41)
NETL
NSTL
万方数据