Our study showed that the ratio of positive VEGF expression in GBM was 63.1% (137/217). This result is similar to that of a previous study (72.5%) [16], but lower than that found in another study (95.8%) [9]. This discrepancy may be due to the following reasons. First, not all gliomas show positive VEGF expression [9,13,16]. Second, the sample size was variable, which may affect the calculation of VEGF expression. In our study, the sample size was 217, larger than that of 24 cases of high-grade glioma in the study conducted by Liu et al. [9]. Third, we defined positive VEGF expression when the positive cell ratio was ≥ 6%. This is because VEGF expression is also positive in healthy brain tissues [13,21].
In our study, two tasks were independently performed using seven radiomic features. The first task aimed to predict positive and negative VEGF expression levels. Of the seven features, three were from T1CE and the remaining four were from T2WI images. The second task was performed to predict high and low-level VEGF expression using another group of seven radiomic features. In this group of features, four out of seven radiomic features were obtained from the data combined by T1CE and T2WI, and the remaining three were from a single MRI sequence. This indicates that the radiomic features from combined data obtained by more than one MRI sequence could be more valuable for radiomic analysis. Although radiogenomic analysis of VEGF expression in diffuse gliomas has been performed previously, the radiomic features used in this study were taken from only one MRI sequence (preoperative T2WI) [14]; radiomic features resulting from combined data with more than one MRI sequence have not been used.
Radiogenomics studies offer a powerful way to characterise the genomic and molecular characteristics of gliomas noninvasively, using semantic or radiomic features. Radiomic features have further been used to facilitate research on their association with molecular markers, clinical outcomes, and other endpoints [22,23]. To date, the biological significance of these individual radiomic features remains unknown [24]. In our study, four out of the seven radiomic features showing significant differences in high and low-level VEGF expression were obtained from the T2WI image data minus the T1CE data set. Pathologically, the T2WI segmented data was composed of GBM itself and a peritumoral high signal area and the T1CE segmented data was the GBM itself. Furthermore, the peritumoral high T2WI area was composed of simultaneous GBM infiltration and vasogenic oedema [25]. The results of our study indicate that for radiomic analysis, combining the raw data according to pathophysiological knowledge could be helpful in obtaining useful radiomic features.
Our analysis used the XGBoost classifier, a fast implementation of the gradient boosting classifier, which has been demonstrated to have excellent scalability and a high running speed [26]. This provides robust prediction results through an iterative process of prediction summation in decision trees fitting to the residual error of the prior ensemble [27] and takes a boosted approach to the decision tree, sequentially adding levels [28]. Compared with the other four algorithms (support vector machine, random forest, naive Bayes, and logistic regression algorithms), this model showed the highest accuracy in ovarian cancer diagnosis [29]. As for gliomas, it improved the classification of MGMT promoter methylation status in IDH1 wildtype GBM [30], efficiently classified transcriptome subtypes in GBM patients from MRI [31], and predicted IDH1 mutation status in gliomas [32]. In our study, this method classified different VEGF expression statuses in GBM with good results.
Exploring VEGF expression in GBM is necessary not only for tumour pathophysiological mechanisms, but also for clinical use. Pathologically, angiogenesis is an important event in GBM owing to its direct influence on tumour invasiveness and poor prognosis [33]. VEGF is a key factor in this event [34]. Clinically, VEGF has been considered as a treatment target for high-grade gliomas, especially GBM [35–38]. Therefore, it is helpful to predict VEGF expression status before anti-VEGF treatment of these tumours. Anti-VEGF treatment is not suitable for GBM patients with negative VEGF expression. For those with positive VEGF expression, tumours with high VEGF expression were more suitable than those with low expression, because this treatment would be more effective in this type of tumour. From this point of view, our research is helpful in prospectively selecting patients with GBM for anti-VEGF treatment.
Our study has some limitations. First, the current study was limited by its retrospective nature, and a prospective study with more patients will be conducted to confirm our findings in future. Second, the whole tumor was segmented on post-contrast axial T1WI, which is an approximate reflection of the lesion. And the different parts of the lesion, such as necrosis and cystic change, the solid component with enhancement, and those without enhancement were not analysed, respectively. Further analysis for those different parts of GBM is needed in future.
In conclusion, our study showed that machine learning analysis using the mRMR algorithm and XGBoost classifier could predict VEGF expression in GBM. Radiomic features from data combinations of different MRI sequences could be more helpful for analysis.