The Role of Heterogeneity Analysis for Differential Diagnosis in Diffusion-Weighted Images of Meningioma Brain Tumors
Mojtaba Safari1, Anahita Fathi Kazerooni1,2, Maryam Babaie3, Mahnaz Nabil4, Mahsa Rostamie1, Parvin Ghavami1, Morteza Saneie Taheri3, and Hamidreza Saligheh Rad1,2

1Quantitative MR Imaging and Spectroscopy Group (QMISG), Research Center for Molecular and Cellular Imaging (RCMCI), Tehran University of Medical Sciences, Tehran, Iran, 2Department of Medical Physics and Biomedical Engineering, School of Medicine, Tehran University of Medical Sciences, Tehran, Iran, 3Radiology Department, School of Medicine, Shahid Beheshti University of Medical Sciences, Tehran, Iran, 4Department of Mathematics, Islamic Azad University, Qazvin Branch, Qazvin, Iran

Synopsis

Meningioma brain tumors constitute the majority of adult primary brain tumors, in which the role of apparent diffusion coefficient (ADC) is controversial. We hypothesize that analysis of the heterogeneity within a tumorous ecological region can reveal biological tissue properties, which could further assist decision making about the optimum patient-specific treatment strategy. In the present work, we propose an automated computer-aided diagnosis method for phenotyping meningioma brain tumors, based on features representing spatial heterogeneity in ADC-maps, with classification accuracy of 85.1%. In conclusion, it is demonstrated that heterogeneity of meningioma brain tumors can be a potential discriminating biomarker of tumor malignancy.

Purpose

Accurate phenotyping of meningioma brain tumors, as the most dominant primary form of brain tumors, is crucial for patient-specific surgical and treatment planning. There remains controversy about the role of diffusion-weighted imaging (DWI) and the extracted apparent diffusion coefficient (ADC)-map in characterization of meningioma brain tumors. Some studies have reported significantly lower ADC values in malignant meningiomas than benign cases¹, but these results were not confirmed by other studies². We hypothesize that spatial analysis of heterogeneity within tumorous area on ADC-maps, which represents the ecological environment created by tumor population, could reveal various biological properties of the tumor, based on water content and cellular density. This can further provide potential biomarkers of tumor malignancy for clinical decision support.

Materials and Methods

We retrospectively examined 44 patients (36 women, 9 men; age range, 11-80 years; mean age, 52 years) who had undergone biopsy or surgical resection of the tumor and histopathological diagnosis. All examinations were performed on a 1.5T MR scanner (Siemens Avanto). The MRI protocol consisted of axial 2D T1-weighted image acquired before and after injection of contrast agent with TR/TE = 420/9 ms, slice thickness = 5.5 mm, flip angle = 90°, field of view (FOV) = 240×240 mm2, matrix size = 256×256; and DW imaging acquired using SE-EPI sequence with TR/TE = 3300/109 ms, slice thickness = 5.5 mm, flip angle = 90°, No. of averages = 4, FOV = 250×250 mm2; matrix size = 128×128, spacing between slices = 7.15 mm, b-values= 50, 1000 s/mm². ADC-maps were directly calculated from DW images on the MR console. The overall procedure consisted of five main steps: (1) Co-registration: The ADC-maps were automatically co-registered with their counterpart post-contrast T1-w (T1C) images through rigid intra-subject registration using FSL software package (http://fsl.fmrib.ox.ac.uk/fsl/). (2) Mask Creation: The tumor border was delineated by selecting regions of interest (ROIs) on T1C images, to create a mask of tumor. (3) Feature Extraction: The tumor mask was overlaid on the ADC-maps to investigate heterogeneity within the region. Several spatial features consisting of (a) first-order histogram measures, (b) higher-order textural methods, including gray-level co-occurrence matrix (GLCM) and run-length matrix (RLM) features, and (c) Gabor features, were computed to explore their significance in discrimination of tumor types and determine the most accurate features for automated classification of meningioma brain tumors. (4) Feature Selection: To decrement the feature space complexity, forward selection Akaike information criterion (AIC) feature selection method was employed³. (5) Classification: Support vector machine classifier with radial basis functions (SVM-RBF) was employed for classification of benign and malignant tumor groups based on the selected features.

Results and Conclusions

SVM-RBF classifier was trained on 30% of the data (with 100 times random selection) for devising the best predictive model of meningioma brain tumor malignancy and was tested for the accuracy on the remaining 70% of the data in each of the 100 iterations. Cross-validation was performed using leave-one-out method and the classification performance was evaluated and reported in terms of sensitivity, specificity, and accuracy (Table 1). It was indicated that after feature selection by AIC-forward technique, the selected features were majorly those of GLCM and RLM textural features, which are representative of heterogeneity of the tumorous region. Fig. 1 shows the classification accuracy attained with increasing number of retained features for the proposed classification scheme. It can be inferred from the plot that the fluctuations of accuracy are small and the method is not very sensitive to the number of selected features; so, we can use small number of features to reduce the computation time and complexity of classification without significantly decreasing the accuracy. The feature selection method can eliminate redundant features, reduce the noise and build groupings that are both robust and accurate. In conclusion, it was shown that the heterogeneity of cellular density and distribution within meningioma tumors on ADC-maps, could be discriminating biomarker of tumor malignancy. This was indicated by inspecting the feature selection strategy, where among all retained features, textural features representing the spatial heterogeneity within a tumorous region were predominantly present as you can see in Fig. 1. This implies that if efficient features of tumor properties are extracted for differentiation of brain tumors, high accuracy could be achieved with fewer number of features. In this study, we thrived to develop an accurate computer-assisted diagnosis (CAD) method for differential diagnosis of benign and malignant meningioma brain tumors from ADC-maps.

Acknowledgements

No acknowledgement found.

References

References: [1] Filippi, Christopher G., et al. "Appearance of meningiomas on diffusion-weighted images: correlating diffusion constants with histopathologic findings." American journal of neuroradiology 22.1 (2001): 65-72.[2] Kawahara, Yosuke, et al. "Prediction of high-grade meningioma by preoperative MRI assessment." Journal of neuro-oncology 108.1 (2012): 147-152. [3] Posada, David, and Thomas R. Buckley. "Model selection and model averaging in phylogenetics: advantages of Akaike information criterion and Bayesian approaches over likelihood ratio tests." Systematic biology 53.5 (2004): 793-808.

Figures

Table 1. The result of SVM-RBF classification for malignant and benign meningioma brain tumors with and without feature selection strategies.

Fig 1. Top: Classification accuracy for discriminating benign and malignant meningioma brain tumors vs. the number of retained features. Below: The number of features selected employing feature extraction methods.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
2811