MRI radiomics combined with ResNet101 deep learning for differentiating lumbar spine brucella spondylitis and spinal metastases
10.13929/j.issn.1003-3289.2025.06.023
- VernacularTitle:MRI影像组学联合ResNet101深度学习鉴别腰椎布鲁氏菌性脊柱炎与脊柱转移癌
- Author:
Yupu LI
1
;
Pengfei ZHAO
;
Xiaojuan ZHANG
;
Zhaojing ZHANG
;
Ziyi WANG
;
Pengfei QIAO
Author Information
1. 内蒙古医科大学第一临床医学院,内蒙古呼和浩特 010059;阳泉市第一人民医院医学影像科,山西阳泉 045000
- Publication Type:Journal Article
- Keywords:
brucellosis;
spondylitis;
spinal cord neoplasms;
neoplasm metastasis;
deep learning;
radiomics;
magnetic resonance imaging
- From:
Chinese Journal of Medical Imaging Technology
2025;41(6):958-962
- CountryChina
- Language:Chinese
-
Abstract:
Objective To observe the value of MRI radiomics combined with ResNet101 deep learning for differentiating lumbar spine brucella spondylitis(BS)and spinal metastases(SM).Methods Seventy-one cases of lumbar spine BS and the same amount of lumbar spine SM patients were retrospectively enrolled in training set,while 33 cases of lumbar spine BS and the same amount of lumbar spine SM patients were enrolled in test set.Clinical features were screened with univariate and multivariate logistic analysis,and a clinical model(Mclinic)was constructed.ROI of lesions were drawn on lumbar sagittal T 2WI,then radiomics features were extracted to construct a radiomics model(Mradiomics).ResNet101 deep learning was integrated with radiomics,then deep learning radiomics features were extracted to construct deep learning radiomics model(MDL+R).Finally a combined model(Mcombined)was constructed through combining clinical features and deep learning radiomics features.The efficacy of the above models for differentiating BS and SM were analyzed.Results Significant differences of patients' age and proportion of fever and accessory involvement were found between BS and SM patients in training and test sets(all P＜0.05),and univariate and multivariate logistic analysis showed the latter two were clinical features(both P＜0.001).The area under the curve(AUC)of Mclinic for differentiating lumbar spine BS and SM was 0.794 and 0.773 in training set and test set,of Mradiomics was 0.895 and 0.791,of MDL+R was 0.926 and 0.882,while of Mcombined was 0.967 and 0.906,respectively.AUC of Mcombined was the highest in training set(all P＜0.05),while in test set,AUC of Mcombined was significantly higher than that of Mclinic and Mradiomics(both P＜0.05).Conclusion MRI radiomics combined with ResNet101 deep learning was helpful for differentiating lumbar spine BS and SM.Combining with clinical data could improve its diagnostic efficacy.
