In this study, we have developed and validated a MRI-Based radiomics nomogram for the differentiation of cervical spine ORN from metastasis in patients with NPC after RT. We found that the radiomics nomogram showed good calibration and discrimination, with an AUC of 0.725 in training set and 0.720 in validation set, respectively. Our results indicated that MRI-Based radiomics may be used as a noninvasive tool for differentiating cervical spine ORN from metastasis after RT.
NPC is one of highly invasive and metastatic head and neck cancer, and cervical spine ORN is a serious complication in NPC after RT [4, 5]. Accurately diagnose cervical spine ORN and distinguish it from bone metastasis is quite important, because an improper diagnosis may create excessive and hurtful chemoradiotherapy for patients. Recently, MRI has been recommended as a very useful technique for identification of benign and malignant vertebral diseases [32, 33]. There are several studies have clarified the value of MRI for diagnosis of ORN, displayed that cervical spine ORN could be misdiagnosed as bone metastasis, because cervical ORN could show soft-tissue masses and present abnormal enhancement [6, 9, 10].
Recently, radiomics features have shown great prospect in identification of malignant and benign bone marrow diseases, including differentiation of primarily malignant and benign bone tumors [12], discrimination of benign and malignant vertebral compression fractures [25], and differentiation metastatic and completely responded sclerotic bone lesion [24]. Particularly, a recent study demonstrated that MRI-based radiomics features could be used to assess the early structural change of femoral head after RT and may showed potential value to predict RT-induced femoral fractures [23]. However, the value of radiomics features in the characterization of ORN and metastasis is still unclear. In this study, we found eight radiomics features based on CE-T1WI were significantly associated with the differentiation of cervical spine ORN and metastasis. Meanwhile, in line with previous studies performed in other field [12, 26, 28], we found these discriminative features that selected to differentiate ORN and metastasis were most frequently derived from GLCM categories.
In the present study, we firstly explored the performance of MRI-based radiomics nomogram for differentiation of cervical spine ORN from metastasis. We found that the radiomics nomogram constructed in training set showed good discrimination efficiency, with an AUC value of 0.725, sensitivity of 84.3% and specificity of 61.4%. Then, we verified the value of this radiomics nomogram in the validation set and still showed good discrimination, with an AUC of 0.720, sensitivity of 80.0%, specificity of 64.0%. Thus, MRI-based radiomics may be a non-invasive imaging biomarker for differentiating cervical spine ORN from metastasis after RT. The results supported previous studies in which MRI-based radiomics could be applied to differentiate vertebral metastasis from benign lesions [25, 34].
The major issue for the clinical application of the nomogram is based on the need to interpret individual net benefits. Nevertheless, the discrimination efficiency and calibration may not acquire the clinical consequences of a particular level of discrimination or degree of miscalibration [35, 36]. To address this issue, we assessed the clinical use of the nomogram by using decision curve analysis (DCA) in the combined training and validation set. This new strategy offers insight into clinical outcomes based on the threshold probability, from which the net benefit could be obtained [14, 28]. In this study, DCA indicated that if the threshold probability of a lesion for diagnosis as ORN is > 12%, in this cases, using the radiomics nomogram to diagnose ORN adds net benefit than either the treat-all-patients scheme or the treat-none scheme. In line with a previous study showed that the threshold probability was > 10% for prediction of lymph node metastasis in colorectal cancer [15].
Our study had some limitations. First, this was a retrospective study performed in single center with relatively small sample size. Thus, multicenter validation is needed to achieve strong evidence for its clinical application. Second, as described in previous studies [6, 9, 10], pathologic confirmation for cervical spine ORN and bone metastasis was not available attributed to the relatively high risks related to biopsy of cervical spine (eg, injury to the vertebral artery or the cervical spinal cord). Third, only radiomics features are selected to construct nomogram model, because the object of this study was cervical spine lesion, considering the fact that patients could involve single or multiple lesions, include patients’ clinical factors could create selection bias.
In conclusion, we preliminarily developed and validated a MRI-Based radiomics nomogram for the differentiation of cervical spine ORN from metastasis in patients with NPC after RT. MRI-based radiomics nomogram may serve as a noninvasive visual diagnostic tool for differentiation cervical spine ORN from metastasis.