Abstract：
BACKGROUND: Degenerative cervical myelopathy (DCM) is the most common form of atraumatic spinal cord injury globally. Degeneration of spinal discs, bony osteophyte growth and ligament pathology results in physical compression of the spinal cord contributing to damage of white matter tracts and grey matter cellular populations. This results in an insidious neurological and functional decline in patients which can lead to paralysis. Magnetic resonance imaging (MRI) confirms the diagnosis of DCM and is a prerequisite to surgical intervention, the only known treatment for this disorder. Unfortunately, there is a weak correlation between features of current commonly acquired MRI scans (\"community MRI, cMRI\") and the degree of disability experienced by a patient.
OBJECTIVE: This study examines the predictive ability of current MRI sequences relative to \"advanced MRI\" (aMRI) metrics designed to detect evidence of spinal cord injury secondary to degenerative myelopathy. We hypothesize that the utilization of higher fidelity aMRI scans will increase the effectiveness of machine learning models predicting DCM severity and may ultimately lead to a more efficient protocol for identifying patients in need of surgical intervention.
METHODS: Single institution analysis of imaging registry of patients with DCM.
METHODS: A total of 296 patients in the cMRI group and 228 patients in the aMRI group.
METHODS: Physiologic measures: accuracy of machine learning algorithms to detect severity of DCM assessed clinically based on the modified Japanese Orthopedic Association (mJOA) scale.
METHODS: Patients enrolled in the Canadian Spine Outcomes Research Network registry with DCM were screened and 296 cervical spine MRIs acquired in cMRI were compared with 228 aMRI acquisitions. aMRI acquisitions consisted of diffusion tensor imaging, magnetization transfer, T2-weighted, and T2*-weighted images. The cMRI group consisted of only T2-weighted MRI scans. Various machine learning models were applied to both MRI groups to assess accuracy of prediction of baseline disease severity assessed clinically using the mJOA scale for cervical myelopathy.
RESULTS: Through the utilization of Random Forest Classifiers, disease severity was predicted with 41.8% accuracy in cMRI scans and 73.3% in the aMRI scans. Across different predictive model variations tested, the aMRI scans consistently produced higher prediction accuracies compared to the cMRI counterparts.
CONCLUSIONS: aMRI metrics perform better in machine learning models at predicting disease severity of patients with DCM. Continued work is needed to refine these models and address DCM severity class imbalance concerns, ultimately improving model confidence for clinical implementation.
OBJECTIVE: This study examines the predictive ability of current MRI sequences relative to \"advanced MRI\" (aMRI) metrics designed to detect evidence of spinal cord injury secondary to degenerative myelopathy. We hypothesize that the utilization of higher fidelity aMRI scans will increase the effectiveness of machine learning models predicting DCM severity and may ultimately lead to a more efficient protocol for identifying patients in need of surgical intervention.
METHODS: Single institution analysis of imaging registry of patients with DCM.
METHODS: A total of 296 patients in the cMRI group and 228 patients in the aMRI group.
METHODS: Physiologic measures: accuracy of machine learning algorithms to detect severity of DCM assessed clinically based on the modified Japanese Orthopedic Association (mJOA) scale.
METHODS: Patients enrolled in the Canadian Spine Outcomes Research Network registry with DCM were screened and 296 cervical spine MRIs acquired in cMRI were compared with 228 aMRI acquisitions. aMRI acquisitions consisted of diffusion tensor imaging, magnetization transfer, T2-weighted, and T2*-weighted images. The cMRI group consisted of only T2-weighted MRI scans. Various machine learning models were applied to both MRI groups to assess accuracy of prediction of baseline disease severity assessed clinically using the mJOA scale for cervical myelopathy.
RESULTS: Through the utilization of Random Forest Classifiers, disease severity was predicted with 41.8% accuracy in cMRI scans and 73.3% in the aMRI scans. Across different predictive model variations tested, the aMRI scans consistently produced higher prediction accuracies compared to the cMRI counterparts.
CONCLUSIONS: aMRI metrics perform better in machine learning models at predicting disease severity of patients with DCM. Continued work is needed to refine these models and address DCM severity class imbalance concerns, ultimately improving model confidence for clinical implementation.
摘要：
背景：退行性脊髓型颈椎病（DCM）是全球范围内最常见的无创伤脊髓损伤形式。椎间盘退变,骨性骨赘生长和韧带病理导致脊髓的物理压迫，导致白质束和灰质细胞群的损伤。这导致患者的潜在的神经和功能下降，这可能导致瘫痪。磁共振成像（MRI）证实了DCM的诊断，是手术干预的先决条件，唯一已知的治疗这种疾病的方法。不幸的是,当前常见的MRI扫描特征之间存在弱相关性(“社区MRI，cMRI\“)和患者经历的残疾程度。
目的：本研究检查了当前MRI序列相对于“高级MRI”（aMRI）指标的预测能力，该指标旨在检测退行性脊髓病继发脊髓损伤的证据。我们假设利用更高保真度的aMRI扫描将提高机器学习模型预测DCM严重程度的有效性，并可能最终导致更有效的方案来识别需要手术干预的患者。
方法：DCM患者影像学登记的单机构分析。
方法：cMRI组296例，aMRI组228例。
方法：生理措施：根据改良的日本骨科协会（mJOA）量表，临床评估机器学习算法检测DCM严重程度的准确性。
方法：纳入加拿大脊柱预后研究网络注册的DCM患者进行筛选，并将cMRI中获得的296例颈椎MRI与228例aMRI进行比较。aMRI采集包括扩散张量成像，磁化转移，T2加权，和T2*加权图像。cMRI组仅包括T2加权MRI扫描。将各种机器学习模型应用于两个MRI组，以评估使用mJOA量表对颈椎病进行临床评估的基线疾病严重程度预测的准确性。
结果：通过利用随机森林分类器，在cMRI扫描和aMRI扫描中预测疾病严重程度的准确率分别为41.8%和73.3%.在测试的不同预测模型变体中，与cMRI扫描相比，aMRI扫描始终产生更高的预测准确性。
结论：aMRI指标在预测DCM患者疾病严重程度的机器学习模型中表现更好。需要继续工作来完善这些模型并解决DCM严重等级不平衡问题，最终提高模型对临床实施的信心。
目的：本研究检查了当前MRI序列相对于“高级MRI”（aMRI）指标的预测能力，该指标旨在检测退行性脊髓病继发脊髓损伤的证据。我们假设利用更高保真度的aMRI扫描将提高机器学习模型预测DCM严重程度的有效性，并可能最终导致更有效的方案来识别需要手术干预的患者。
方法：DCM患者影像学登记的单机构分析。
方法：cMRI组296例，aMRI组228例。
方法：生理措施：根据改良的日本骨科协会（mJOA）量表，临床评估机器学习算法检测DCM严重程度的准确性。
方法：纳入加拿大脊柱预后研究网络注册的DCM患者进行筛选，并将cMRI中获得的296例颈椎MRI与228例aMRI进行比较。aMRI采集包括扩散张量成像，磁化转移，T2加权，和T2*加权图像。cMRI组仅包括T2加权MRI扫描。将各种机器学习模型应用于两个MRI组，以评估使用mJOA量表对颈椎病进行临床评估的基线疾病严重程度预测的准确性。
结果：通过利用随机森林分类器，在cMRI扫描和aMRI扫描中预测疾病严重程度的准确率分别为41.8%和73.3%.在测试的不同预测模型变体中，与cMRI扫描相比，aMRI扫描始终产生更高的预测准确性。
结论：aMRI指标在预测DCM患者疾病严重程度的机器学习模型中表现更好。需要继续工作来完善这些模型并解决DCM严重等级不平衡问题，最终提高模型对临床实施的信心。
