3080

Radiomics Approach for Prediction of Tumor Recurrence and Progression of Skull Base Meningioma

Ching-Chung Ko¹, Yang Zhang², Jeon Hor Chen^2,3, Peter Chang², Daniel Chow², Tiffany Kwong², and Min-Ying Lydia Su²

¹Section of Neuroradiology, Department of Medical Imaging, Chi-Mei Medical Center, Tainan, Taiwan, ²Department of Radiological Sciences, University of California, Irvine, CA, United States, ³Department of Radiology, E-Da Hospital and I-Shou University, Kaohsiung, Taiwan

Synopsis

A subset of low grade skull base meningiomas (SBM) shows early progression/recurrence (P/R). In clinical practice, one of the main challenges in the treatment of SBM is to determine factors that correlate with P/R. This study investigated the role of radiomics for the prediction of P/R. Sixty patients diagnosed with benign SBM were studied. Totally 99 descriptors were extracted from the various MR sequences. The prediction accuracy of P/R was 90% and the AUC of the prediction model was 0.94. Our study also noted that subsequent P/R of SBM after surgery was not associated with the completeness of tumor resection.

Background and Purpose

Approximate 20–30% of meningiomas grow in the skull base [1]. Although the majority of skull base meningiomas (SBM) are low grade (WHO grade I), a subset of WHO grade I SBM shows early progression/recurrence (P/R) in the first year after surgical resection [2]. Because of the complex anatomic structures in skull base, complete surgical resection of the tumor (Simpson Grade I-III resection) is often difficult to achieve safely [3]. SBMs are often associated with involvement of complex neurovascular structures and brainstem compared with lesions not arising from the skull base, conservative follow up or subtotal tumor resection (STR) to avoid subsequent surgical complications are also the treatment of choice [4]. In clinical practice, one of the main challenges in the treatment of SBM is to determine factors that correlate with P/R. The conventional qualitative MR imaging findings such as mushroom shape, bone osteolysis, dural tail, and proximity to major sinuses had been reported as the important variables related to P/R [2]. Recently, the association of apparent diffusion coefficient (ADC) and P/R has been reported by our group [5]. Besides ADC, it is very likely that other quantitative imaging parameters may also be associated with tumor recurrence. Thus, in this study, we further investigated the role of radiomics for the prediction of P/R.

Materials and Methods

Sixty patients (age 26–75 years; median age, 56 years) included in this study were diagnosed with low grade (WHO grade I) SBM by MRI and pathological confirmation. The median follow-up time was 41 months (range 12-115 months), and a total of 21 patients were found to have recurrence or progression, and 39 patients remained disease-free or without any sign of progression. The MRI images were acquired using a 1.5T or a 3.0T scanner. The protocols of MR imaging included axial and sagittal spin echo T1-weighted imaging (T1WI), fast spin-echo T2-weighted imaging (T2WI), fluid attenuated inversion recovery (FLAIR), T2*-weighted gradient-recalled echo (GRE), contrast-enhanced (CE) T1WI in axial and coronal sections, and diffusion weighted imaging (DWI). Figure 1 shows the flowchart of the analysis process. The lesion was segmented from subtracted contrast enhancement images. For each lesion, the operator placed an initial region of interest (ROI) indicating the lesion location, and also decided the beginning and ending slices that contained the lesion. Then the outline of the lesion ROI on each imaging slice was automatically obtained using the fuzzy c-means (FCM) clustering based algorithm [6]. The ROIs from all imaging slices containing this lesion were combined to obtain 3D information of the whole lesion. Then 3D connected-component labeling was applied to remove scattered voxels not connecting to the main lesion ROI, and hole-filling to include all voxels contained within the main ROI which were labeled as non-lesion. The segmented tumor mask was co-registered to T2W images and ADC map to localize the tumor location on corresponding images using affine transformation. This process was done by FLIRT. Within segmented tumor on enhanced T1W images, T2W images and ADC maps, 13 histogram features and 20 textural GLCM features were extracted on each modality [7]. Thus, totally we obtained 99 descriptors. To evaluate the importance of these features in differentiate patients with and without recurrence, random forest algorithms were utilized via Bootstrap-aggregated decision trees [8]. Three features, including T1 Max Probability, T1 Cluster Shade, ADC Correlation, with the highest importance were selected to build a decision tree with 5 leaves that makes coarse distinction between classes. The total number of split was 4 (Figure 2). This procedure was implemented in Matlab 2018b.

Results

Of the 60 patients, 37 had complete tumor resection (Simpson Grade I-III) and 23 had incomplete tumor resection (Simpson Grade IV and V). Figure 3 shows two patients with Simpson Grade 2 with complete tumor resection in surgery: one had recurrence and the other did not in subsequent follow-up. Figure 4 shows another two patients with Simpson Grade 4 with incomplete tumor resection in surgery: one had progression and the other did not in subsequent follow-up. Figure 5 shows box plot for three features which were extracted from random forest algorithm and included in the final prediction model. The final classification results showed 18 true positive cases, 36 true negative cases, 3 false positive cases, and 3 false negative cases. The overall prediction accuracy is 90% and the AUC of the prediction model is 0.94. We also analyzed the impact of tumor resection completeness based on Simpson Grades, and found that it was not a risk factor associated with the progression/recurrence.

Discussion

This study attempted to use radiomics approach in limited cases of skull base meningioma for the prediction of tumor recurrence or progression after surgery. Our results showed that, with total 99 descriptors of histogram features and textural GLCM features extracted from segmented tumor on enhanced T1W images, T2W images and ADC maps, a prediction accuracy of 90% and an AUC of 0.94 of the prediction model were achieved. The results were better than our previous work merely using ADC, which was measured by operator-defined ROIs, as the predictor [5].

Acknowledgements

This study was supported in part by NIH R01 CA127927.

References

[1] Mansouri A, et al. Surgically resected skull base meningiomas demonstrate a divergent postoperative recurrence pattern compared with non–skull base meningiomas. J Neurosurg 2016;125:431–40.; [2] Ildan F, et al. Predicting the probability of meningioma recurrence in the preoperative and early postoperative period: a multivariate analysis in the midterm follow-up. Skull Base. 2007;17:157–71.; [3] Simpson D. The recurrence of intracranial meningiomas after surgical treatment J Neurol Neurosurg Psychiatr. 1957;20:22–39.; [4] Sekhar LN, et al. Skull base meningiomas: aggressive resection. Neurosurgery. 2015; 62 Suppl 1: 30–49.; [5]. Ko CC, et al. Applications of diffusion-weighted MR imaging in brain tumors. J Neurooncol. 2018;138(1):63-71.; [6] Nie K, et al. Quantitative analysis of lesion morphology and texture features for diagnostic prediction in breast MRI. Academic Radiology. 2008;15(12):1513-25.; [7] Haralick RM, et al. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics. 1973(6):610-21.; [8] Segal MR. Machine learning benchmarks and random forest regression. Center for Bioinformatics & Molecular Biostatistics. 2004.

Figures

Figure 1. Flowchart of the analysis procedures. The tumor is first segmented based on T1-weighted contrast enhancement image, and then the tumor ROI is mapped to T2-weighted image and ADC map. On each of the three images, a total of 20 GLCM texture and 13 histogram-based parameter are extracted, and the total of 99 parameters for each case are used to develop the classification model to differentiate between patients without progression/recurrence vs. those with progression/recurrence.

Figure 2. Random forest algorithms via Bootstrap-aggregated decision trees to differentiate between 39 patients without and 21 with progression/recurrence. Three parameters, including T1_Max Probability, ADC_Correlation, and T1_Cluster Shade, are selected to build the final classification model. The results based on the selected threshold show that of 39 cases without P/R, 36 are corrected predicted (true negative) and three are wrong (false positive). Of 21 patients with P/R, 18 patients are correctly predicted (true positive) and three are wrong (false negative).

Figure 3. Two cases with Simpson Grade 2 complete resection in surgery. From left to right: the ADC map, T1-weighted contrast-enhanced image, and T2-weighted images are shown. The upper panel is a 56 y/o female with a meningioma in the fontal base, who does not have any sign of recurrence in the subsequent follow-up. The lower panel is a 43 y/o female with a meningioma in the right frontal base, who shows recurrence in the subsequent follow-up after one year.

Figure 4. Two cases with Simpson Grade 4 with incomplete resection in surgery. From left to right: the ADC map, T1-weighted contrast-enhanced image, and T2-weighted images are shown. The upper panel is a 73 y/o female with a meningioma in the cerebellar area, who does not have any sign of progression in the subsequent follow-up. The lower panel is a 34 y/o female with a meningioma in the left tentorial area, who shows progression in the subsequent follow-up after one year.

Figure 5. The box plot of three imaging features (T1_Max Probability, T1_Cluster Shade and ADC_Correlation) selected in the random forest algorithm to differentiate patients without Progression/Recurrence and those show sign of Progression/Recurrence.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)

3080