“Prediction of Glioma Grades Based on Radiomic Features Inside and Outside Tumors Using Multiparametric MRI Images”

Research Background

Glioma is the most common primary brain tumor in the central nervous system, accounting for 80% of adult malignant brain tumors. In clinical practice, treatment decisions often require individualized adjustments based on the grade of the tumor. The World Health Organization (WHO) classifies gliomas into four grades (I-IV), further categorizing them into low-grade gliomas (LGG, Grades I and II) and high-grade gliomas (HGG, Grades III and IV). Accurate grading of gliomas is crucial for developing treatment plans, implementing personalized therapy, and predicting prognosis and survival time. Currently, the diagnosis of glioma grades mainly relies on surgical biopsy or histopathological analysis. However, this diagnostic method is invasive and in some cases not suitable for patients, highlighting the need for a non-invasive and highly accurate glioma grading system.

Magnetic Resonance Imaging (MRI) has become a popular non-invasive tool for radiologists to diagnose brain tumors over the past few years. Although experienced radiologists can easily detect tumors from MRI sequences with the naked eye, identifying glioma grades remains challenging due to tumor heterogeneity. In recent years, significant progress has been made in the grading of gliomas using preoperative multiparametric MRI (mpMRI) scans and radiomic features extracted from these scans. For example, Zacharaki et al. developed a computer-aided diagnosis (CAD) system to distinguish brain tumor types using Support Vector Machine-Recursive Feature Elimination (SVM-RFE) method. Vamvakas et al. also adopted the SVM-RFE algorithm to select 21 radiomic features from mpMRI data to predict glioma grades, achieving good performance on a small sample set. Despite these studies’ relative success in distinguishing HGG from LGG, extracting more valuable radiomic features to improve prediction accuracy remains a challenge.

Research Source

This study was conducted by Jianhong Cheng, Jin Liu, Hailin Yue from the School of Computer Science and Engineering at Central South University, in collaboration with renowned medical scholars Harrison Bai and Yi Pan. The paper was published in the journal “IEEE/ACM Transactions on Computational Biology and Bioinformatics.” The aim of this study is to propose a non-invasive and highly accurate method for glioma grading by combining features extracted both inside (Intratumoral volume, ITV) and outside (Peritumoral volume, PTV) the tumor from preoperative mpMRI scans.

Research Workflow

The entire study includes several steps:

Data and Preprocessing

The study data comes from the BRATS challenge dataset, including 285 preoperative mpMRI scans (210 HGG and 75 LGG), and 65 preoperative mpMRI scans reorganized from the BRATS2019 dataset (48 HGG and 17 LGG) as an external validation dataset. Each subject has four scanning modes: T1-weighted (T1), enhanced T1-weighted (T1gd), T2-weighted (T2), and T2 Fluid-Attenuated Inversion Recovery (FLAIR). All scan data were first reoriented to the same coordinate system, resampled to 1mm isotropic voxel size, and the skull was removed using a brain extraction tool.

Volume Definition and Segmentation

According to clinical studies, BRATS organizers identified three Volumes of Interest (VOIs): the enhancing portion of the tumor core (ET), necrotic and non-enhancing portions (NET), and peritumoral edema (PED). Non-enhancing tumors contain only non-enhancing tumor structures. To further investigate the impact of PTV on glioma grading, we developed a method to capture volumes within a 1mm to maximum 5mm radius outside the tumor core.

Radiomic Feature Extraction

Using the PyRadiomics toolbox, radiomic features were extracted from each VOI, processed using various filtering methods and classification techniques to quantify seven categories of features: first-order statistics-based features, 2D and 3D shape features, Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Neighboring Gray Tone Difference Matrix (NGTDM), and Gray Level Dependence Matrix (GLDM). A total of 2153 quantitative radiomic features were extracted from each VOI.

Feature Selection and Classifier Modeling

The LASSO regression method was used to select the most important features from each mode, and then the best features from all modes were combined. The Minimum Redundancy Maximum Relevance (MRMR) algorithm was used to eliminate redundant features. The top-ranked radiomic features were input into classifiers to construct a radiomic signature for predicting glioma grades, evaluated using five-fold cross-validation.

Research Results

We conducted technical validation on the BRATS2017 dataset and compared our results with those of other similar studies. The results showed that the non-enhancing tumor region and the 1mm region around the tumor had high predictive performance. By combining features from these two regions, the classifier’s predictive ability was further enhanced, reaching a maximum AUC value of 0.975.

Results on the external validation dataset further confirmed our method’s strong generalization performance, achieving high accuracy and sensitivity, demonstrating the significant application value of the proposed radiomic features in distinguishing glioma grades.

Research Significance

This study proposes a novel non-invasive and highly accurate method for glioma grading based on intra-tumoral and peri-tumoral features from preoperative mpMRI scans. By comprehensively utilizing the rich data from multiparametric MRI images, this study provides new imaging biomarkers for clinical diagnosis, potentially supporting personalized treatment and prognosis prediction.

Highlights and Innovations

• A novel method combining radiomic features from both inside and outside the tumor has been proposed, improving the accuracy of glioma grade prediction.
• An innovative algorithm accurately capturing the fuzzy boundary regions outside the tumor has been designed and embedded into the PyRadiomics toolbox.
• By integrating multiparametric MRI imaging and various machine learning classifiers, the potential value of multi-dimensional features has been deeply explored.

Conclusion

The comprehensive method combining radiomic features from inside and outside the tumor shows excellent performance in predicting glioma grades, indicating its great potential and application prospects in computer-aided diagnosis. Further research and empirical validation of imaging biomarkers will contribute to more comprehensive tumor evaluation and treatment decision-making.