In the present study,
radiomic analysis on a 3T mp-MRI dataset was performed for glioma grade classification between low- and high-grade tumors,
demonstrating 95.5% Accuracy and 95.5% AUC in predicting glioma grades,
utilizing 21 mp-MRI radiomic features.
The justification for implementing the specific SVM feature selection and classification methods,
is based on the predictive robustness indicated by similar studies regarding glioma grading in the past.
In a computer-aided-diagnostic approach Chen et al.
[14] utilized SVM-RFE for selecting textural features,
derived from CNN-based segments of conventional MRI data,
and XGBoost classification presenting 91,27% accuracy.
Citak-Er et al.
[8] have proposed a sophisticated SVM-RFE implementation, with different tumor ROIs mean values of an mp-MR dataset.
Subsequently,
they utilized the SVM-RFE outcome for training a linear SVM classifier with 93% classification accuracy.
Our study exploiting a comparable patient sample size and MR sequences confirms the increased diagnostic ability provided by SVM-RFE and linear SVM classification of mp-MRI data.
In addition,
the higher accuracy value presented in our study demonstrates the potential role of utilizing radiomic features.
Tian et al.
in a recent study [9],
which was based on the same group’s initial research [13],
proposed a mpMRI glioma grading classification scheme based on SVM-RFE and RBF kernelized SVM,
showing 96% accuracy and 98% AUC values.
Even though their study presents slightly better performance compared to our study,
our model is achieving comparable results utilizing a smaller number of radiomic features (21 vs.
28),
increasing its relative efficiency.
Since the two studies follow similar feature extraction,
selection and classification methodologies,
it is obvious that our study’s good performance should be attributed to the addition of Diffusion Tensor and Spectroscopic imaging data utilized.
As shown in Table 1,
an important number of features extracted from these techniques,
have shown high discriminative ability,
as this is already investigated and confirmed [15].
More specifically, previous studies have reported the value of textural features of Apparent Diffusion Coefficient (ADC) as potential biomarkers for glioma grade differentiation [16].
In a recent study,
Raja et al.
[17] investigated the contribution of Diffusion Tensor and Diffusion Kurtosis Imaging in gliomas grading.
Their texture-based features have shown significant differences regarding several DTI parameters,
except for FA,
which comes in agreement with the results of our study.
However,
we have found that the anisotropic diffusion tensor component,
as expressed by the pure anisotropy (q) has proven to be of great importance (5 out of the 21 features were derived for q maps) and could play an essential role in glioma heterogeneity assessment.
In addition,
MR spectroscopy is a powerful technique for evaluating brain tumor metabolic processes with an increased diagnostic impact. Previous studies support the potential of MRS metabolic ratios in brain pathology differentiation [18],
especially when combined with other advanced techniques.
Even though a statistically significant difference for the specific ratios was not observed in our study,
which might be expected since we are comparing gliomas,
however the MRS derived Lipids/Cr ratio was the highest ranked feature.
Consequently,
the lipids concentration in glioma’s tumor core,
which is proportional to the extent of tumor’s necrotic component,
may serve as a robust imaging biomarker in differentiating between Low- and High-grade gliomas.
In conclusion,
the recent technological advancements in the field of medical imaging have given rise to the incorporation of innovative methodologies regarding tumor phenotypic characteristics quantification and multiparametric data analysis which aid in improving the clinical decision support.
The current study presents a comprehensive methodological perspective for evaluating MRI derived phenotypic characteristics for glioma grading,
based on multiparametric MR neuroimaging data and state-of-the-art radiomic analysis methods.
It shows that radiomic features derived from mp-MRI could be used for accurate glioma classification by exploiting the underlying pathophysiology.