In this study, we evaluated the performance of qualitative, quantitative parameters and radiomics models based on T1WI and T2WI to discriminate intracranial SFTs and ATMs. Among the qualitative and quantitative parameters on T1WI and T2WI, our results showed that hyperintensity on T1WI and hypointensity on T2WI were valuable predicted features for differentiating intracranial SFTs with ATMs. Among the inter-imaging model comparisons, it demonstrated that radiomics-based models’ performance tended to be better than those of qualitative and quantitative parameters-based models, especially when using the T1&T2WI joint model. In both the training and testing sets, the AUCs of the radiomics models reached 0.908–0.961 and 0.854–0.942, respectively. These results show that the radiomics process has unearthed valuable information to distinguish intracranial SFTs and ATMs. In addition, qualitative parameters on T1WI and T2WI, and quantitative parameters on T1WI were negatively correlated with CVF among intracranial SFTs.
Many researchers have explored the correlation between clinical-radiological features and the differentiation of intracranial SFTs and ATMs, such as age, gender ratio, “dural tail” sign, bone erosion, and intratumoral flow void [14, 23–26]. In our study, age and gender distribution presented statistical differences in intracranial SFTs and ATMs, in concordance with previous studies. Some previous studies demonstrated that intracranial SFTs were generally isointense with the gray matter on T1WI and T2WI, with no significant differences with ATMs [26, 27]. However, He et al. reported that signal value difference between white matter and tumor on T1WI or T2WI could distinguish SFTs and angiomatous meningiomas [17]. But the applicability of the signal value difference is limited on account of affecting by field strength and MRI machine types. In the current study, a qualitative assessment of a five-point scale was evaluated on T1WI or T2WI, and they were helpful for the determination of intracranial SFTs and ATMs. Significantly, the diagnostic confidence for SFTs may be improved, when the five-point scale was 4–5 and 1–3 on T1WI and T2WI, respectively. In a small case study by Weon et al., 4 of 6 lesions showed high or isointensity on T1WI and low signal intensity on T2WI, in line with our results, while another 2 lesions showed low signal intensity on both T1WI and T2WI [28]. Previous studies supposed that hypointensity on T2WI was corresponding with collagenous hypocellular regions [29]. However, the reason for the area presenting as high signal intensity on T1WI among these cases has not been known. The AUCs of qualitative parameters on T1WI were 0.743 in the training set, and 0.913 in the testing set. We think that this heterogeneity between the training and testing sets is attributed to the potential bias caused by random allocation of patients and the small sample size. In addition, quantitative parameters were also evaluated in our study. But only quantitative parameters on T1WI (rT1thalamus and rT1centrum semiovale) presented statistically different distribution in intracranial SFTs and ATMs, and lower diagnostic performance than qualitative parameters on T1WI. Overall, the qualitative parameters on T1WI and T2WI, and quantitative parameters on T1WI can be applied for the prediction of intracranial SFTs, with the advantages of easy achievement.
Radiomics can be used to mine high-throughout imaging features from non-invasive sequences to improve cancer diagnosis. To our knowledge, the association between radiomics features and differentiation of intracranial SFTs and ATMs has not been evaluated. However, previous studies have applied the radiomics models based on T1WI or T2WI to distinguish between SFTs and meningiomas, especially angiomatous meningiomas, achieved AUCs ranging from 0.762 to 0.918 [20, 21, 30]. In our sample cohort, we applied the LR method to construct three radiomics models (T1WI, T2WI, and T1&T2WI joint models) for non-invasive preoperative discrimination of the two entities, which also achieved good performance (AUC = 0.854–0.942). Regarding radiomics features, our results demonstrated that the majority of selected features turned out to be texture features, which are important markers for intratumoral homogeneity. SFTs are usually presented as heterogenous signal intensity, attributing to two distinct components-collagenous and hypercellular regions within the lesions [31]. Most interesting, the elongation and sphericity features were also chosen to construct the radiomics models. These shape features may reflect growth patterns. Recently, Yan et al. found that tumor shape features were useful indicators in grading meningiomas, similar to our study [32]. From the histological perspective, intracranial SFTs are vascular-rich tumors and commonly grow more rapidly than meningiomas, and may, thus, result in irregular and lobular shape [33]. Among these radiomics models, the T1&T2WI joint model provided better predictive performance than T1WI or T2WI model on its own. The results of our study indicated that radiomics features from both T1WI and T2WI sequences could reflect the heterogeneity of tumors comprehensively. Moreover, the radiomics models outperformed most of the radiologists’ qualitative and quantitative models in the training sets. However, the discrimination ability was not found between radiomics models and the radiologist’s assessment and in testing sets. We speculated that this may be on account of the small number of the testing data. Nevertheless, we still assume that radiomics models may have potential value to assist doctors in determining SFTs and ATMs, especially for less experienced radiologists.
Based on the significant difference in qualitative, quantitative, and radiomics features between SFTs and ATMs in our study, we assume that some histopathological features may be attributed to imaging results. As we know, SFTs encompassed low cellularity embedded in abundant collagen bundles and highly cellular components with slit-like vascular spaces [1, 34], correlating with “ying-yang” sign on T2WI. A recent study in a small number of orbital SFTs also indicated that kurtosis in T2WI negatively associated with collagen content [22]. To our knowledge, the correlation between imaging features and CVF among intracranial SFTs has not been reported. Interestingly, our results indicated that SFTs with lower signal intensity on T2WI and T1WI represented more collagen within tumors. This result may be used to quantitatively evaluate the “ying-yang” sign on T2WI from histological respect, in accordance with previous reports [31]. Indeed, although the imaging findings of hyperintensity on T1WI in intracranial SFTs have been reported previously, and could not also be explained by collagen in our study, we still assumed it as a useful imaging biomarker in the discrimination of intracranial SFTs and ATMs.
Limitations
First, potential selection biases intrinsic to this retrospective single-center study existed. Second, the number of patients with SFTs was a relatively small. However, considering that SFTs were relatively rare types of intracranial tumors, the number of patients in the current study was acceptable. Third, although the results based on non-enhanced T1WI and T2WI are promising in our study cohorts for the discrimination of SFTs and ATMs, enhanced sequences might provide extra information, which was not available for all patients in the current study. Lastly, CVF was only calculated on SFTs for evaluating the correlation with imaging features, given that SFTs have the characteristics of abundant collagen rather than meningiomas.