The present study established an accurate classifier to distinguish parasellar cavernous hemangiomas from meningiomas by integrating a large panel of radiomic features. An efficient classifier was obtained by comparing five MRI sequences from 1.5 T and 3.0 T MR scanners at three medical imaging centers, bolstering its generalizability. The MRI-based radiomic classifier outperformed the neuroradiologists in terms of diagnostic accuracy, sensitivity, and specificity.
Radiomics can provide additional metabolic and biological information in addition to the traditional MRI metrics. Gray contrast, uniformity, depth, and texture roughness have been used to study tumor grading, prediction of genomic information, and differentiation of lesion and non-lesion images25–27. The present study found that higher-order features could better reflect the degree of tumor heterogeneity and texture information. A GLSZM can quantify gray-level zones in an image to reflect tumor heterogeneity at a local scale. The coefficient of High Gray-Level Zone Emphasis was the largest, which measured the distribution of the higher gray-level values. Larger values indicated a larger proportion of high gray-level values and size zones in the image28. Tumor heterogeneity usually reflected the gray contrast variation of the image. Therefore, the GLSZM was more sensitive in distinguishing parasellar cavernous hemangiomas from meningiomas.
The present study focused on the inadequacy of visual examination in differentiating parasellar cavernous hemangiomas from meningiomas in order to assess the clinical role of radiomics in facilitating and enhancing visual analysis by radiologists. T2WI and ADC sequences had a good practical value in constructing radiomics models. T2WI was characterized by a high signal-to-noise ratio, homogeneity, and significant hyperintensity16,29. The radiomics model constructed based on T2WI had a high diagnostic accuracy and stability in distinguishing parasellar hemangiomas and meningiomas. Radiomics models based on ADC maps were equally valuable in the differential diagnosis when parasellar hemangiomas and meningiomas were identified. However, the detection rate of lesions in ADC maps was about 78.7%. Contrary to our general view, the accuracy of the radiomics model based on CE-T1WI was low, although it was improved in different ways. This might be influenced by different types of cavernous hemangiomas and meningiomas30–32.
The present study compared the detection rates for parasellar cavernous hemangiomas and meningiomas obtained using different MRI sequences, which has not been attempted in previous studies. The detection rate of DWI and ADC maps was 78.7% (37/47), with a cut-off diameter of 2.25 cm. The detection rate of T2WI, T1WI, and CE-T1WI was 100%, which was more conducive to the establishment of radiomics models.
Different classifier algorithms may lead to different results. The present results suggested that the radiomics models combined with SVM and KNN classifiers had better diagnostic performance in distinguishing between parasellar cavernous hemangiomas and meningiomas. SVM has been proposed by Cortes et al. in 1995 as a binary classifier based on supervised learning33,34. The critical concept of SVM involves the use of a hyperplane to define decision boundaries to separate different classes of data points. This technique finds support vectors with a high discrimination and maximizes the interval between classes. It has good adaptability and discrimination ability. The K-nearest neighbor (KNN) method is mostly used for image classification. This object classification is based on the distance between its neighbors and is mainly used to solve regression and classification problems. By selecting the KNN points of a sample when the nearest neighbors belong to a certain category, the sample is determined to belong to that category. Several previous studies have demonstrated KNN's excellent and stable performance using different datasets, which was similar to the present result35–37. Consistent with our study, other classifiers also suffer from over-fitting. This is manifested by the fact that the training set is too accurate, while the validation set cannot achieve the expected ideal results. In addition, there are too many feature dimensions, parameters, and noise, which lead to a too-perfect prediction of the fitted function in the training set. However, the prediction results in the new data test set were low. In the present study, SVM and KNN classifiers were suggested for use as radiological diagnostic models to distinguish between parasellar cavernous hemangiomas and meningiomas.
There are several limitations in the present study. First, the sample size was relatively small and needs to be further explored. Second, different types of parasellar cavernous hemangiomas and meningiomas were not considered. Third, the differential diagnosis mainly focused on parasellar hemangiomas and meningiomas. Other parasellar tumors that are relatively easy to diagnose were not included in the study.
In conclusion, the proposed T2WI-based radiomics model combining SVM and KNN classifiers showed favorable predictive efficacy in the preoperative differential diagnosis between parasellar cavernous hemangiomas and meningiomas. It had more general applicability in complementing conventional imaging modalities and as an alternative to functional imaging. Moreover, the more readily available T2WI could provide higher detection rates and more texture features. Other imaging modalities based on T2WI for differentiating parasellar cavernous hemangiomas and meningiomas need to be explored.