International Journal of Radiation Oncology biology physics www.redjournal.org Imaging in Radiation Oncology Clinical Applications of Quantitative 3-Dimensional MRI Analysis for Pediatric Embryonal Brain Tumors Jared H. Hara, BS,* Ashley Wu, BS,* Javier E. Villanueva-Meyer, MD,y Gilmer Valdes, PhD,* Vikas Daggubati, BS,* Sabine Mueller, MD, PhD,z Timothy D. Solberg, PhD,* Steve E. Braunstein, MD, PhD,* Olivier Morin, PhD,* and David R. Raleigh, MD, PhD*,z Departments of *Radiation Oncology, yRadiology, and zNeurological Surgery, University of California, San Francisco, San Francisco, California Received Dec 19, 2017, and in revised form May 22, 2018. Accepted for publication May 28, 2018. Summary The relationship between clinical outcomes and imaging characteristics of pediatric embryonal brain tumors is poorly understood. We analyzed embryonal brain tumor magnetic resonance images to identify quantitative radiomic features associated with clinical outcomes. Our results demonstrate that features quantifying primary tumor size and heterogeneity are associated with patient age, neuraxis metastases, histology, and recurrence. These data suggest that 3dimensional magnetic Purpose: To investigate the prognostic utility of quantitative 3-dimensional magnetic resonance imaging radiomic analysis for primary pediatric embryonal brain tumors. Methods and Materials: Thirty-four pediatric patients with embryonal brain tumor with concurrent preoperative T1-weighted postcontrast (T1PG) and T2-weighted fluid-attenuated inversion recovery (FLAIR) magnetic resonance images were identified from an institutional database. The median follow-up period was 5.2 years. Radiomic features were extracted from axial T1PG and FLAIR contours using MATLAB, and 15 features were selected for analysis based on qualitative radiographic features with prognostic significance for pediatric embryonal brain tumors. Logistic regression, linear regression, receiver operating characteristic curves, the Harrell C index, and the Somer D index were used to test the relationships between radiomic features and demographic variables, as well as clinical outcomes. Results: Pediatric embryonal brain tumors in older patients had an increased normalized mean tumor intensity (P Z .05, T1PG), decreased tumor volume (P Z .02, T1PG), and increased markers of heterogeneity (P .01, T1PG and FLAIR) relative to those in younger patients. We identified 10 quantitative radiomic features that delineated medulloblastoma, pineoblastoma, and supratentorial primitive neuroectodermal tumor, including size and heterogeneity (P .05, T1PG and FLAIR). Decreased markers of tumor heterogeneity were predictive of neuraxis metastases and trended toward significance (P Z .1, FLAIR). Tumors with an increased size (area under the Reprint requests to: David R. Raleigh, MD, PhD, Departments of Radiation Oncology and Neurological Surgery, University of California, San Francisco, 505 Parnassus Ave, L08/L75, Box 0226, San Francisco, CA 94143. Tel: (415) 353-3675; E-mail: email@example.com Int J Radiation Oncol Biol Phys, Vol. -, No. -, pp. 1e13, 2018 0360-3016/$ - see front matter Published by Elsevier Inc. https://doi.org/10.1016/j.ijrobp.2018.05.077 O.M. and D.R.R. contributed equally to this work. Conflict of interest: none. Supplementary material for this article can www.redjournal.org. be found at 2 International Journal of Radiation Oncology Biology Physics Hara et al. resonance imaging analysis has the potential to identify radiomic risk features for pediatric patients with embryonal brain tumors. curve Z 0.7, FLAIR) and decreased heterogeneity (area under the curve Z 0.7, FLAIR) at diagnosis were more likely to recur. Conclusions: Quantitative radiomic features are associated with pediatric embryonal brain tumor patient age, histology, neuraxis metastases, and recurrence. These data suggest that quantitative 3-dimensional magnetic resonance imaging radiomic analysis has the potential to identify radiomic risk features for pediatric patients with embryonal brain tumors. Published by Elsevier Inc. Introduction Embryonal brain tumors, such as medulloblastoma, pineoblastoma, and supratentorial primitive neuroectodermal tumor (sPNET), are the most common pediatric brain tumors. These highly malignant neoplasms are often diagnosed in young children, in whom curative treatments involving surgery, radiation therapy, and chemotherapy are often associated with significant long-term side effects (1-3). Thus, pediatric embryonal brain tumor treatments are frequently individualized according to clinical prognostic features to maximize tumor control and minimize toxicity (4-8). In that regard, molecular subgrouping has provided key clinical prognostic information that has guided several prospective de-escalation trials for Wnt-associated medulloblastoma (6-8). However, medulloblastoma and other pediatric embryonal brain tumors are heterogeneous, and there is an unmet need for novel strategies to guide adjuvant therapy (9-11). Recently, several studies have shown that physician-defined, qualitative radiographic features are associated with molecular subgroups of pediatric embryonal brain tumors and may provide useful information to adapt adjuvant therapy (12-16). Building on those successes, the purpose of this study was to evaluate the prognostic utility of quantitative radiomic features for pediatric embryonal brain tumors. Many studies have identified correlations between imaging characteristics and clinical or molecular prognostic features in diverse malignancies such as lung cancer, head and neck cancer, and glioblastoma (17-22). In embryonal brain tumors, size and qualitative imaging features such as tumor location, cystic tumor morphology, hemorrhage, and intratumoral necrosis are associated with distinct molecular features (13, 23-25). Magnetic resonance imaging (MRI) with gadolinium enhancement differs between adult and pediatric medulloblastoma patients (26-34), in whom there are also strong associations between MRI characteristics, overall survival (12, 16), and neuraxis metastasis (35, 36). More recently, studies have revealed similar correlations among qualitative magnetic resonance spectroscopy features, medulloblastoma location, and molecular subgroup (14-16, 37, 38). By using quantitative imaging features, radiomics can function as a guide that may provide novel information beyond what a physician can consciously observe (39). For pediatric embryonal brain tumors, physician-defined imaging features have been discovered to correlate with some but not all important genetic and prognostic characteristics. Thus, we hypothesized that quantitative radiomic analysis of features derived from unbiased tumor masks might enhance prognostication for pediatric embryonal brain tumors. In support of this hypothesis, quantitative textural analysis can be used to discriminate between different pediatric brain tumor histologies (40-43), but these approaches have not been previously reported for pediatric embryonal brain tumors. We explored the prognostic value and complementary role of morphologic, normalized intensity and textural radiomic features to (1) differentiate between pediatric embryonal brain tumor histologies preoperatively, (2) identify tumors with neuraxis metastases at presentation, and (3) identify patients who are at risk of recurrence. The specific objectives of this study were to (1) extract 3-dimensional (3D) radiomic features from common MRI sequences (T1-weighted postcontrast [T1PG] and T2weighted fluid-attenuated inversion recovery [FLAIR]) and (2) define a framework for radiomics in pediatric embryonal brain tumors by identifying relationships among individual 3D MRI features, patient demographic characteristics, and clinical outcomes. Our results reveal associations between quantitative imaging features of pediatric embryonal brain tumors and patient age, tumor histology, neuraxis metastasis, and recurrence, suggesting that 3D radiomic analysis could be used as a noninvasive method for risk stratification in pediatric patients with embryonal brain tumor. Methods and Materials Study design and patient population We performed a retrospective institutional database search for embryonal brain tumors diagnosed between June 1988 and October 2014 with institutional review board approval (protocol 14-15432) and a waiver of consent. The study population consisted of diverse patients treated over several therapeutic eras; some patients were enrolled on clinical trials, but the majority were treated according to eraspecific trial parameters. Only patients with concurrent preoperative T1PG and FLAIR MRI sequences from a 1.5or 3-T scanner were included (Tables E1 and E2, available Volume - Number - 2018 Table 1 Pediatric embryonal brain tumor radiomics Patient, tumor, and treatment characteristics Data Patients and tumors Total patients Age at diagnosis (range), y Sex Male Female M status M0* Mþ* Histology Medulloblastoma Morphology Desmoplastic Nodular or desmoplastic Classic Anaplastic Not otherwise specified Molecular mutation Wnt Sonic hedgehog Myc 17pq deletion Unknown Pineoblastoma sPNET Treatment Gross total resection Subtotal resection Radiation therapy CSI Local dose, Gy CSI dose, Gy Upfront radiation therapy Age at diagnosis, y Time to radiation therapy, d Delayed radiation therapy Age at diagnosis, y Received delayed radiation therapy Time to radiation therapy, d Chemotherapy Total no. of cycles Vincristine Vincristine and carboplatin CCNU, vincristine, and cisplatin Thiotepa, cyclophosphamide, etoposide, and vincristine Vincristine, cisplatin, and cyclophosphamide Carboplatin and etoposide Carboplatin and etoposide alternating with ifosfamide and etoposide Cisplatin and cyclophosphamide Etoposide, cyclophosphamide, and vincristine Vincristine, cisplatin, cyclophosphamide, and cis-retinoic acid Cisplatin, cyclophosphamide, vincristine, and etoposide Vincristine, cisplatin, cyclophosphamide or mesna, etoposide, and methotrexate Cisplatin, vincristine, CCNU, and cyclophosphamide Stem cell transplant Shunt placement Salvage therapy after recurrence 34 6.9 (IQR, 4.1-13.0) 13 (38.2%) 21 (61.7%) 30 (88.2%) 4 (11.8%) 22 (64.7%) 3 1 2 7 21 (8.8%) (2.9%) (5.9%) (20.6%) (61.8%) 1 1 1 3 28 6 (17.6%) 6 (17.6%) 27 7 29 27 55.8 23.4 27 7.1 26 7 2.7 2 172 32 5 (79.4%) (20.6%) (85.2%) (79.4%) (IQR, 54.0-55.8) (IQR, 23.4-36.0) (79.4%) (IQR, 5.1-13.6) (IQR, 22-33) (20.6%) (IQR, 2.5-5.4) (5.9%) (IQR, 161-183) (94.1%) (IQR, 4-8) 3 5 5 3 6 1 1 1 1 1 2 2 1 6 (17.6%) 2 (6%) 9 (26.5%) (continued on next page) 3 4 Hara et al. International Journal of Radiation Oncology Biology Physics Table 1 (continued ) Data Surgery Radiation therapy Chemotherapy Stem cell transplant Outcome Median follow-up, y Recurrencey Local failure Regional and distant failure (Mþ) Recurrent regional or distant failure New area of failure Median time to recurrence, y Death Median time to death, y 2 6 7 4 5.15 9 4 5 2 3 2.2 6 2.31 (5.9%) (17.6%) (20.6%) (11.8%) (IQR, 2.69-7.99) (26.5%) (44.4%) (55.5%) (40.0%) (60.0%) (95% CI, 1.5-3.0) (17.6%) (95% CI, 1.4-5.6) Abbreviations: CCNU Z lomustine; CI Z confidence interval; CSI Z craniospinal irradiation; IQR Z interquartile range; sPNET Z supratentorial primitive neuroectodermal tumor. * The extent of metastatic spread was defined as Mþ for M1 to M4 disease after cerebrospinal fluid cytologic analysis and spinal magnetic resonance imaging for staging, consistent with the Chang criteria (45). y Recurrence was defined as any local (resection cavity), regional, or distant metastatic failure using the Chang criteria (45). M0 was defined as localized disease at initial therapy; Mþ, disseminated disease at initial therapy; local failure, resection cavity failure; and regional and distant failure, neuraxial or distant visceral organ metastases after therapy. online at www.redjournal.org) (44). The oldest scan year included in the study was 2001. Patient demographic, pathologic, and clinical risk factors are summarized in Table 1. Patient age was defined as the age at initial diagnosis. The extent of metastatic spread was defined as Mþ for M1 to M4 disease after cerebrospinal fluid (CSF) cytologic analysis and spinal MRI for staging, consistent with the criteria of Chang et al (45). The extent of resection was based on postoperative MRI and defined as gross total resection (GTR) or subtotal resection (1.5 cm2). For the purposes of this study, patients with near-total resection were grouped with GTR patients (46, 47). Histopathologic re-review was performed by neuropathologists according to current diagnostic criteria (48, 49). Recurrence was defined as any local (resection cavity), regional, or distant metastatic failure using the Chang criteria (45); patients with regional or distant metastatic failure were further subclassified according to growth of neuraxial metastases that were present at initial presentation versus emergence of new areas of disease. Survival was defined as the length of time from diagnosis to the date of last follow-up, to the date of radiographic progression for progression-free survival, or to the date of death for overall survival (Fig. 1). Radiomic feature selection and extraction Preoperative images were exported, and regions of interest (ROIs) were delineated on both T1PG and FLAIR axial imaging sequences using MIM Vista (version 6.4.9; MIM Software, Cleveland, OH) by a single radiation oncologist (D.R.R) to limit interreviewer variation. Both tumor and control ROIs were contoured (Fig. 2A), the latter of which were delineated at gray-matter junctions using a 5-mm brush on 3 consecutive slices from both T1PG and FLAIR axial images for low-intensity normalization. ROIs were then expanded and contracted serially by 1 mm using an automated MIM workflow (Fig. 2B), anonymized, and exported to MATLAB 2017a (The MathWorks, Natick, MA) for feature extraction via an open-source development radiomic package called the Quantitative Imaging Prediction Modeling platform (https://github.com/qipm/QIPM) (50, 51). This radiomic package has been validated by the Image Biomarker Standardization Initiative. After image quality analysis, 4 FLAIR sequences were excluded because of software incompatibility. Although the Quantitative Imaging Prediction Modeling platform is capable of computing thousands of features, the number of features was kept close to the number of patients to specifically limit the chance of false-positive discoveries (52). Accordingly, quantitative imaging features were selected for further analysis after physicist (O.M.) and neuroradiologist (J.E.V.) review of radiomic parameters that tractably quantify qualitative radiographic features with known prognostic value for pediatric embryonal brain tumors (as discussed in the “Introduction” section; Table 2). Features selected included (1) morphologic features, which capture tumor size and shape; (2) statistical features, which calculate nondiscretized intensity values within tumors; and (3) textural features, which express how combinations of gray levels are spatially distributed, connected, or summed (Table E3, available online at www.redjournal.org). The gray-level co-occurrence and size zone matrix features were selected to quantify heterogeneity and the amount of Volume - Number - 2018 B 100 75 Overall survival (%) Progression free survival (%) A Pediatric embryonal brain tumor radiomics 50 25 0 5 10 Time (years) 15 75 50 25 20 0 5 10 Time (years) 15 20 Progression-free survival (A) and overall survival (B) for patients with embryonal brain tumor included in the study. necrotic tissue and/or brain invasion. Normalized intensity in statistical features was selected with the intention to quantify hypointense, isointense, and hyperintense radiographic presentation. This approach generated a total of 35 radiomic A 100 0 0 Fig. 1. 5 Medulloblastoma Pineoblastoma sPNET features per imaging study that were computed to quantify size, shape, normalized intensity, and texture of each tumor. Normalized intensity was defined as the ROI mean signal of the tumor divided by the ROI mean signal of a gray-matter Control B T2-Fluid Attenuated Inversion Recovery Merge T2-Fluid Attenuated Inversion Recovery T1-Post Contrast T1-Post Contrast Fig. 2. Representative magnetic resonance imaging scans of pediatric embryonal brain tumors in patients. The 3 histologic tumor types (axial T1-weighted postcontrast [T1PG] [top and bottom rows] and axial T2-weighted fluid-attenuated inversion recovery [FLAIR] [middle row] magnetic resonance images) (A) and contour expansions for image feature extractions (B) are shown. (A) Supratentorial primitive neuroectodermal tumors (sPNETs) were noted to have markedly larger tumor volumes, whereas pineoblastomas were associated with the smallest maximal 3-dimensional tumor diameter. Pineoblastomas were noted to have increased markers of heterogeneity relative to medulloblastomas and sPNETs. Tumors contoured from T1PG images are shown in magenta, and tumors contoured from T2-weighted FLAIR images are shown in cyan. (B) Representative axial T1PG (left) and axial T2-weighted FLAIR (right) images are shown for radiomic feature computation on 1-mm expansions and contractions of the original regions of interest to add robustness in feature selection. 6 International Journal of Radiation Oncology Biology Physics Hara et al. Table 2 Associations between radiographic and radiomic features with relevance to pediatric embryonal brain tumors Radiographic feature Clinical association Radiomic feature Description Size sPNET is often very large (62) Fmorph_volume Fmorph_diam Enhancement Primary pediatric embryonal brain tumors that are hypointense on T1 (22, 27), are hyperintense on T2 (23, 24, 45), and show T2 contrast enhancement are often associated with leptomeningeal spread (26) Associated with medulloblastoma (21, 24, 29) Medulloblastoma has high cellular density; association of high density with low signal on T1-weighted imaging (22, 47) Heterogeneous appearing medulloblastomas are associated with an improved prognosis (63), and sPNET is heterogeneous appearing (62); medulloblastoma is associated with necrosis and calcifications (21, 24), whereas both sPNET and medulloblastoma are associated with hemorrhage (24, 27, 30) Fstat_mean_norm Fstat_var Morphologic volume Morphologic maximum 3-dimensional diameter Nondiscretized normalized mean intensity Nondiscretized intensity variance Fmorph_sphericity Fmorph_comp1 Morphologic sphericity Morphologic compactness Fcm_contrast Fcm_dissimilarity Fszm_zs_var Fszm_zs_entr Fngt_contrast Fngt_busyness Fngt_complexity Fngt_strength Gray-level co-occurrence matrix contrast Gray-level co-occurrence matrix dissimilarity Gray-level size zone matrix variance Gray-level size zone matrix entropy Gray-level co-occurrence matrix contrast Neighborhood gray tone busyness Neighborhood gray tone complexity Neighborhood gray tone strength Cystic change Cell density Heterogeneity Necrosis Hemorrhage Calcification Abbreviation: sPNET Z supratentorial primitive neuroectodermal tumor. control region. Radiomic features were also computed on the 1-mm expansion and contraction of the original ROI to add robustness in feature selection. Features were calculated using a 3-mm isotropic interpolation for nontextural and textural features, linear interpolation for ROI mask, and a fixed bin number of 32 for discretization of the primary image intensity. Among the initial 35 features calculated, we selected the features with the largest observed variance in the cohort. This feature selection eliminated features that were nearly constant for the cohort studied. Ultimately, 15 radiomic features were selected for further analyses based on correlations with physician-defined, qualitative radiographic characteristics with known prognostic value for pediatric embryonal tumors. The radiomic features selected encompassed primary tumor size, enhancement, heterogeneity (hemorrhagic, calcifications, necrosis), and cystic components (Table 2), and textural features were used as markers to quantify heterogeneity. Statistical analysis All statistical analyses were performed using Stata Statistical Software (version 15; StataCorp, College Station, TX) and JMP (version 13.0; SAS Institute, Cary, NC). Progression-free survival and overall survival were analyzed by the KaplanMeier method. We computed z scores for each radiomic feature to transform the data to comparable scales for regression analyses. The ability of individual radiomic features to distinguish between patients by clinical outcomes was assessed by performing logistic regression for sex and M status, multinomial logistic regression for histology, linear regression for age, and Cox regression for recurrence and survival outcomes. P .05 was considered statistically significant (Table 3; Table E4, available online at www. redjournal.org). Receiver operating characteristic curves were built to evaluate the adequacy of fitted logistic regression models, as conveyed by area under the curve (AUC) values. Asymptotic confidence intervals (CIs) were constructed to provide measures of uncertainty around the curves for out-of-sample predictions, and AUC values 0.7 were considered statistically significant (Table 3; Table E4, available online at www.redjournal.org). Harrell C indexes were computed to evaluate the predictive power of radiomic features in fitted Cox regression models, whereas Somer D indexes were computed to estimate rank parameters with respect to censored survival time. Indexes and respective 95% CIs were reported (Table 3; Table E4, available online at www.redjournal.org) (53-56). Individual radiomic features and clinically relevant predictor variables of age, GTR, and M status were included in multivariate Cox regression analyses for recurrence and survival outcomes. Hazard ratios were reported to reveal each covariate’s predictive power (Tables E6-E18, available online at www.redjournal.org). Spearman correlation heat maps were built to illustrate homogeneity between selected radiomic features (Figs. E1 and E2, available online at www.redjournal.org). Volume - Number - 2018 Results and Discussion Patient characteristics, treatments, and clinical outcomes Thirty-four pediatric embryonal brain tumor patients with both T1PG and FLAIR preoperative imaging were identified (Table 1). The median age at diagnosis was 6.9 years (interquartile range [IQR], 4.1-13.0 years). The median follow-up period was 5.2 years (IQR, 2.7-8.0 years), and the median overall survival period was 2.3 years (95% CI, 1.4-5.6 years). With respect to tumor histology, 22 patients (64.7%) had medulloblastoma, 6 (17.6%) had pineoblastoma, and 6 (17.6%) had sPNET. Four patients (11.8%) had neuraxis or distant metastasis at presentation. Twenty-seven patients (79.4%) underwent GTR and 7 (20.6%) underwent subtotal resection. Before preoperative imaging, 2 patients (6%) underwent shunt placement. Adjuvant radiation therapy was administered in 29 patients (85.2%), and 27 patients received craniospinal irradiation (79.4%), with a median tumor bed dose of 55.8 Gy (IQR, 54.0-55.8 Gy) and median craniospinal dose of 23.4 Gy (IQR, 23.4-36.0 Gy). Upfront radiation therapy was administered in 27 patients (79.4%), and 2 patients (2.9%) were treated with delayed radiation therapy. Adjuvant chemotherapy was used in 32 patients (94.1%), with a median of 5 cycles (IQR, 4-8 cycles), and 6 patients (17.6%) underwent stem cell transplant. During the followup period, 9 tumors recurred (26.5%), with a median time to recurrence of 2.2 years (95% CI, 1.5-3.0 years), and 6 patients died (17.6%) (Fig. 1). After recurrence, 9 patients (26.5%) received salvage treatment (Table 1). Patient age Tumor size and qualitative radiographic metrics of enhancement and heterogeneity are associated with pediatric embryonal brain tumor patient age (26, 30, 57). Our morphologic quantitative findings similarly revealed that tumor volume (P Z .03, T1PG) and maximum 3D diameter (P Z .02, T1PG) were distinctly smaller with increased age (Table 3, Fig. 3A). Likewise, we observed that tumor enhancement intensity, as measured by normalized mean intensity, increased with increasing patient age on T1PG (P Z .05), which was again consistent with qualitative evidence (57). We observed that 4 markers of heterogeneity of embryonal brain tumors increased in older patients, as evidenced by increased gray-level co-occurrence matrix contrast (P Z .003, T1PG; P Z .01, FLAIR) and dissimilarity (P Z .002, T1PG), as well as neighboring gray tone contrast (P Z .003, T1PG) and complexity (P Z .001, T1PG). Of note, 2 of the markers of heterogeneity decreased with age as shown by neighboring gray tone busyness (P Z .03, T1PG) and gray-level size zone matrix variance (P Z .02, T1PG). In addition, markers of heterogeneity were highly correlated with one another Pediatric embryonal brain tumor radiomics 7 (Figs. E1 and E2, available online at www.redjournal.org). Generally, age prediction on radiomic analysis favored T1PG (Table 3). Overall, these findings support qualitative assessments indicating that embryonal brain tumors are usually more heterogeneous in older patients (57). Moreover, the radiomic differences we observed are consistent with the conception that adult and pediatric embryonal brain tumors have distinct biological characteristics (58). Histology Qualitative radiographic studies have identified distinct characteristics for medulloblastoma, pineoblastoma, and sPNET histologies (6, 16, 59-61). Broadly, medulloblastomas are typically hypointense on T1 with heterogeneous contrast enhancement, isointense to hypointense on FLAIR with surrounding edema, and hyperintense with reduced diffusion on diffusion-weighted imaging (DWI) (6, 16, 59). Pineoblastomas tend to appear as large irregular masses with a cystic or necrotic component that are isointense to hypointense on T1 with heterogeneous enhancement and isointense on T2 with restricted diffusion owing to high cellular packing on DWI (61). Finally, sPNETs are variable on T1 with heterogeneous contrast enhancement, a high solid signal on T2 with cystic components, and a markedly low level of peritumoral vasogenic edema (60). Preliminary data have suggested that multimodal MRI with DWI and magnetic resonance spectroscopy can be used to identify pediatric central nervous system tumor type at diagnosis and better detect recurrence (62, 63), but with respect to quantitative radiomic features, there are no prior reports of textural analysis among pediatric embryonal brain tumors. Our quantitative radiomic analyses demonstrated that size is a key feature for differentiating pediatric embryonal brain tumors. Not surprisingly, 2 separate measures of tumor size, including volume (P < .03, T1PG; P Z .05, FLAIR) and maximum 3D diameter (P < .03, T1PG; P Z .04, FLAIR) were associated with specific embryonal brain tumor histologies (Table 3; Fig. 2A; Fig. E3, available online at www .redjournal.org). In that regard, we observed that sPNETs (126.3 cm3; IQR, 33.7-167.4 cm3) were significantly larger than either medulloblastomas (34.6 cm3; IQR, 21.2-39.0 cm3) or pineoblastomas (3.8 cm3; IQR, 3.1-4.5 cm3) (Tables E19 and E20, available online at www.redjournal.org). Similarly, pineoblastomas (2.94 cm; IQR, 2.59-3.34 cm) had a significantly smaller maximum 3D diameter than sPNETs (8.92 cm; IQR, 5.70-10.70 cm) or medulloblastomas (6.00 cm; IQR, 5.50-6.33 cm). Clinically, sPNETs are often larger at the time of detection because of their location and lower likelihood of causing hydrocephalus compared with pineoblastomas or medulloblastomas. Of note, tumors were larger on FLAIR than on T1PG for volume (P < .05) (Table 3). For example, medulloblastomas had a higher volume on FLAIR (44.8 cm3; IQR, 30.0-57.9 cm3) than on T1PG (34.6 cm3; IQR, 21.2-39.0 cm3), which is consistent with the surrounding infiltrative disease commonly seen 8 International Journal of Radiation Oncology Biology Physics Hara et al. Table 3 Univariate analysis of imaging biomarkers from primary pediatric embryonal brain tumors for dependent variables of age, histology, M stage, or recurrence Age: P value M status T1PG Morphologic volume Morphologic compactness Morphologic sphericity Morphologic maximum 3D diameter Nondiscretized normalized mean intensity Nondiscretized intensity variance Gray-level co-occurrence matrix contrast Gray-level co-occurrence matrix dissimilarity Gray-level size zone matrix variance Gray-level size zone matrix entropy Neighborhood gray tone coarseness Neighborhood gray tone contrast Neighborhood gray tone busyness Neighborhood gray tone complexity Neighborhood gray tone strength T1PG FLAIR P value .027* .509 .523 .021* .054* .486 .003* .002* .02* .155 .143 .007* .03* .001 .11 .062 .644 .656 .091 .532 .306 .013* .022* .098 .125 .138 .063 .067 .001 .11 .941 .64 .57 .465 .863 .799 .401 .415 .704 .198 .21 .433 .77 .591 .183 AUC (95% CI) 0.69 0.46 0.46 0.64 0.53 0.45 0.59 0.55 0.36 0.66 0.69 0.58 0.68 0.57 0.70 (0.50-0.88)* (0.20-0.72) (0.20-0.72) (0.46-0.83) (0.26-0.80) (0.13-0.78) (0.32-0.85) (0.29-0.82) (0.14-0.59) (0.44-0.88) (0.51-0.87)* (0.31-0.84) (0.48-0.89)* (0.29-0.84) (0.50-0.89)* FLAIR P value .678 .485 .502 .196 .336 .402 .128 .155 .761 .144 .21 .132 .84 .591 .183 AUC (95% CI) 0.74 0.61 0.61 0.73 0.63 0.67 0.73 0.67 0.31 0.74 0.75 0.77 0.75 0.57 0.70 (0.56-0.92)* (0.38-0.85) (0.38-0.84) (0.55-0.91)* (0.32-0.94) (0.43-0.91) (0.55-0.91)* (0.48-0.85)* (0.12-0.50) (0.56-0.92)* (0.58-0.92)* (0.58-0.97)* (0.57-0.93)* (0.42-0.84) (0.56-0.92)* Abbreviations: 3D Z 3-dimensional; AUC Z area under curve; CI Z confidence interval; FLAIR Z T2-weighted fluid-attenuated inversion recovery; MB Z medulloblastoma; PB Z pineoblastoma; sPNET Z supratentorial primitive neuroectodermal tumor; T1PG Z T1-weighted postcontrast. Logistic regression analysis was used for M status. Cox regression was used for recurrence. Linear regression was used to assess age. Multinomial regression was used to assess histology. The area under the receiver operating characteristic curve was used for neuraxis metastases, and the Harrell C statistic was used for recurrence. * Significant P, AUC, or C-index values. around the bulky core of these tumors (Tables E19 and E20, available online at www.redjournal.org) (16). Pediatric embryonal brain tumor histologies were also markedly different with respect to 8 of the 9 textural features (Fig. 2A). Both gray-level co-occurrence features and size zone matrix, which are markers of tumor heterogeneity (64), demonstrated unique features among histologic subgroups. Gray-level co-occurrence matrix features, which express how combinations of discretized gray levels of neighboring pixels or voxels in a 3D volume are distributed along one of the image directions, were distinct for contrast (P Z .03, T1PG; P Z .009, FLAIR) and dissimilarity (P Z .02, T1PG; P Z .01, FLAIR) among medulloblastomas and pineoblastomas. Similarly, size zone matrix, which measures the number of groups of connected voxels with a specific discretized gray level and size, was different for variance (P Z .03, T1PG) and entropy (P Z .01, T1PG). In addition, neighboring gray tone, another marker of heterogeneity, differed among medulloblastomas and pineoblastomas in terms of radiomic coarseness (P Z .005, T1PG; P < .02, FLAIR), contrast (P Z .02, T1PG and FLAIR), busyness (P < .05, T1PG), and strength (P Z .006, T1PG; P Z .02, FLAIR). Pineoblastoma radiomic features generally had higher scores for heterogeneity (contrast, 66.5 [IQR, 36.8-86.3] on T1PG; dissimilarity, 6.37 [IQR, 4.61-7.40] on T1PG) (Tables E19 and E20, available online at www.redjournal.org), which is consistent with the qualitative radiographic literature (65). Conversely, medulloblastoma radiomic features had comparatively less evidence of heterogeneity (contrast, 30.2 [IQR, 27.7-37.1] on T1PG; dissimilarity, 4.11 [IQR, 3.93-4.52] on T1PG) (10). Thus, quantitative radiomics demonstrates that pineoblastoma is the most internally heterogeneous pediatric embryonal brain tumor and that medulloblastoma is the most homogeneous. Metastases Spinal MRI and lumbar puncture for CSF analysis are standard of care for metastatic staging in pediatric patients with embryonal brain tumor (66). However, CSF analysis is associated with a high false-negative rate, and postoperative or postelumbar puncture spinal imaging is associated with a high false-positive rate because of surgical and hematologic debris and reactive change (67, 68). Moreover, up to 50% of patients with pediatric embryonal brain tumor have equivocal spinal MRI studies (69). As a result, CSF analysis and spinal MRI results are often discordant (66). To address this unmet need, qualitative radiographic studies have suggested that neuroimaging may provide an additional source of evidence of neuraxis metastasis (35, 36). Indeed, we observed that T1PG and FLAIR 3D radiomic features of primary pediatric embryonal brain tumors trended toward discriminating between localized (M0) and disseminated (Mþ) disease. For patients with metastases, those with a larger primary tumor size were observed to have a higher likelihood of neuraxis metastases (maximal tumor diameter AUC Z 0.74, FLAIR). In addition, we observed a decrease in primary tumor heterogeneity, as Volume - Number - 2018 Pediatric embryonal brain tumor radiomics 9 Table 3 Univariate analysis of imaging biomarkers from primary pediatric embryonal brain tumors for dependent variables of age, histology, M stage, or recurrence (continued) Histology: P value Recurrence PB vs MB sPNET vs MB T1PG T1PG FLAIR T1PG FLAIR P value .063 .161 .167 .029* .054* .189 .031* .021* .032* .013* .005* .023* .045* .141 .006* .111 .391 .379 .142 .709 .307 .009* .012* .158 .996 .017* .017* .103 .051* .02* .026* .909 .92 .029* .145 .645 .902 .508 .042* .993 .739 .991 .037* .541 .768 .050* .535 .542 .043* .233 .864 .688 .38 .129 .322 .49 .706 .087 .411 .808 .656 .591 .623 .323 .221 .888 .3 .231 .394 .173 .297 .351 .862 .357 .34 FLAIR C index (95% CI) 0.63 0.53 0.53 0.62 0.63 0.52 0.58 0.60 0.59 0.67 0.63 0.59 0.62 0.60 0.62 (0.42-0.84) (0.31-0.74) (0.31-0.74) (0.41-0.82) (0.43-0.83) (0.29-0.74) (0.39-0.78) (0.39-0.82) (0.37-0.81) (0.45-0.89) (0.41-0.85) (0.35-0.83) (0.40-0.84) (0.37-0.84) (0.39-0.85) P value .886 .208 .205 .261 .572 .436 .186 .219 .608 .271 .225 .18 .95 .357 .34 C index (95% CI) 0.71 0.62 0.62 0.71 0.49 0.41 0.57 0.56 0.43 0.67 0.70 0.62 0.71 0.61 0.72 (0.53-0.88)* (0.41-0.83) (0.41-0.83) (0.54-0.89)* (0.25-0.73) (0.21-0.60) (0.37-0.78) (0.36-0.75) (0.2-0.65) (0.45-0.90)* (0.53-0.87)* (0.40-0.84) (0.52-0.90)* (0.39-0.84) (0.54-0.91) measured by textural features such as neighborhood gray tone coarseness (AUC Z 0.7, T1PG and FLAIR) and strength (AUC Z 0.7, T1PG and FLAIR) (Tables 2 and 3, Fig. 3B). Neuraxis metastasis at presentation trended toward decreased primary tumor radiomic metrics of heterogeneity in additional radiomic FLAIR features, such as neighborhood gray-level co-occurrence matrix contrast (P Z .1, AUC Z 0.8, FLAIR) and dissimilarity (AUC Z 0.7, FLAIR), as well as neighborhood gray tone contrast (P Z .1, AUC Z 0.8, FLAIR), which also delineated metastatic tumors on FLAIR but not on T1PG. Consistently, FLAIR characteristics of primary pediatric embryonal brain tumors are thought to have greater potential to identify patients with distant neuraxis metastasis than T1PG imaging features (31). Of note, these results were not statistically significant on logistic regression, likely because of the limited sample size. This limitation notwithstanding, our data suggest that the radiomic features of primary tumors may provide an additional source of information for differentiating between localized and disseminated pediatric embryonal brain tumors. predict prognosis of pediatric embryonal brain tumors (70). From the standpoint of quantitative radiomics, we discovered that patients who have recurrence trended toward having larger tumors at the time of diagnosis (Fig. 3C, left), as denoted by maximum 3D diameter (AUC Z 0.7, FLAIR) and volume (AUC Z 0.7, FLAIR), and were more likely to ultimately develop recurrent disease (Fig. 3C, right). In addition, we observed that decreased heterogeneity, as measured by neighborhood gray tone coarseness (AUC Z 0.7), contrast (AUC Z 0.7), and busyness (AUC Z 0.7), was associated with disease recurrence on FLAIR. Of note, recurrence was also positively associated with size zone matrix entropy (AUC Z 0.7, T1PG), as observed for metastases at presentation. Unfortunately, these patterns were not significant on univariate analysis, likely again because of the limited sample size (Table 3). Nevertheless, our findings are clinically important because 80% of deaths in our patients resulted from progressive disease (Table 1, Fig. 1). Thus, our data suggest that patients with decreased heterogeneity and/or larger tumors on MRI may benefit from more aggressive therapy. Recurrence Limitations and future directions Beyond patient demographic characteristics and clinical features at diagnosis, we explored the role of quantitative 3D MRI analysis to predict clinical outcomes (Tables 2 and 3). Previous groups have demonstrated the value of qualitative MRI features to screen for recurrence and Aside from the retrospective nature of our study, we were principally limited by a small data set, which may have contributed to nonsignificant findings on univariate analysis (Table 3). Our strict inclusion criteria of only pediatric patients with embryonal brain tumor who had preoperative 10 International Journal of Radiation Oncology Biology Physics Hara et al. A 7-year-old 18-year-old B C Merge T2-Fluid Attenuated Inversion Recovery T1-Post Contrast 2-year-old Fig. 3. Representative magnetic resonance imaging scans of medulloblastoma in patients of various ages: axial T1-weighted postcontrast (top and bottom rows) and axial T2-weighted fluid-attenuated inversion recovery (FLAIR) (middle row) magnetic resonance images (A), neuraxis metastases at presentation (B), and tumor recurrence (C). (A) Images of a 2-year-old patient (left), 7-year-old patient (middle), and 18-year-old patient (right) with medulloblastomas, demonstrating a less contrast-enhancing tumor component in the older patient. (B) Images of a 6-year-old patient with medulloblastoma presenting with neuraxis metastases associated with an increased maximal 3-dimensional primary tumor diameter and decreased markers of primary tumor heterogeneity. (C) Axial T2-weighted FLAIR magnetic resonance images of a 4-yearold patient who initially presented with a right frontal supratentorial primitive neuroectodermal tumor (left) and who underwent a subtotal resection and subsequently received high-dose adjuvant chemotherapy (thiotepa and carboplatin) with stem cell transplant (middle); the patient had a local recurrence 26 months after surgery (right). 3D MRI certainly contributed to this limitation. Thus, because of the small sample size and exploratory nature of our study, we did not correct for multiple comparisons and we acknowledge that these findings cannot directly show statistical significance. In addition, given the large number of patients with medulloblastoma in comparison with pineoblastoma or sPNET, we recognize that our results favor findings for medulloblastoma, with a decreased sensitivity for pineoblastoma or sPNET. Moreover, we were unable to use assigned costs to address class imbalance because of the limited sample size, but we recognize this as a potential avenue of future investigation. However, we observed several radiomic features that were associated with key clinical parameters on both T1PG and FLAIR sequences for each of these tumors; thus, we are optimistic that our findings will be generalizable to the larger population of pediatric patients with embryonal brain tumor. The patients included in this study were referred from multiple centers with different MRI scanners, which have inherent differences in image characteristics. To internally account for this variation, the normal brain contour was used to first normalize the brain and then discretize the gray level to account for magnetic resonance scan signal variation. Moreover, we observed no systematic differences in MRI radiomic characteristics when patients were dichotomized according to magnet strength (Tables E1 and E2, available online at www.redjournal.org). However, we observed a difference in radiomic characteristics, specifically for gray-level co-occurrence size zone matrix between magnet strengths, but there was no change in the results after standardizing the gray-level co-occurrence size zone matrix features internally with respect to magnet strength. Recent studies have suggested that interscanner variability does not prohibit the use of a quantitative radiomic analysis but have noted that single-scanner studies may improve predictive quantitative radiomic analysis outcomes. Moreover, these results indicate that similar textural features are enhanced by magnetic resonance images obtained at different hospitals (71). However, more work must be done to further explore the technique to implement multi-institutional transferability of these images. Volume - Number - 2018 We identified significant correlation between tumor size and features for heterogeneity (Figs. E1 and E2, available online at www.redjournal.org). These results suggest that even stronger associations may be identified with a larger data set. With the high degree of correlation within feature classes and between feature size and heterogeneity, a larger data set is needed to assess intraclass correlation and identify underlying components that better characterize correlated features. Because there is currently no established method to select textural features that correlate with specific histopathologic or clinical correlates such as heterogeneity, the results of our study should be taken with caution. However, the features selected were supported mathematically, conceptually, and biologically based on the current understanding of quantitative radiomic features, as well as radiographic correlates, which were consistent with our quantitative radiomic feature analysis. These results support the notion that there is a clinical need and potential for further establishment of radiomic textural features with their clinical correlate. We observed that medulloblastoma, pineoblastoma, and sPNET are associated with markedly different radiomic features, and we cannot rule out the possibility that associations stemmed from histologic differences as opposed to clinical outcomes. We also recognize that genetic subgroups have recently played a role in diagnostic, prognostic, and predictive outcomes of these patients (48). Because there are currently no studies describing the impact of quantitative imaging features on prognosis in the context of multiple histologies and varied risk categories, these exploratory results support the potential for future multicenter studies to further quantify and validate the relationship between clinical outcomes and 3D MRI quantitative features. Conclusions Pediatric embryonal tumor risk stratification is based on clinical and molecular factors. Previous studies have shown that pediatric embryonal tumors present with distinct qualitative imaging characteristics (23-25), but little is known about the relationship between quantitative imaging characteristics and clinical outcomes in this patient population. We performed quantitative 3D T1PG and FLAIR MRI radiomic analyses of pediatric embryonal brain tumors in the context of patient demographic characteristics and clinical outcomes. Our results identified radiomic features that are consistent with known qualitative radiographic features associated with patient age and tumor histology, such as contrast enhancement and markers of tumor heterogeneity. In addition, our results identified potential predictive radiomic features of neuraxis metastasis and recurrence such as tumor size and decreased markers of tumor heterogeneity. Given the limitation of sample size and the limited molecular subgrouping data, these findings should be taken with caution, and future multicenter studies Pediatric embryonal brain tumor radiomics 11 are needed to validate these results. Overall, our data support a potential paradigm to use 3D textural analysis as an objective measurement of qualitative radiographic features and suggest that MRI radiomics has potential to identify radiomic risk features for patients who may benefit from aggressive adjuvant therapy. References 1. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2015. CA Cancer J Clin 2015;65:5-29. 2. Ostrom QT, Gittleman H, Fulop J, et al. CBTRUS statistical report: Primary brain and central nervous system tumors diagnosed in the United States in 2008-2012. Neuro Oncol 2015;17(Suppl. 4):iv1-iv62. 3. Mueller S, Fullerton HJ, Stratton K, et al. Radiation, atherosclerotic risk factors, and stroke risk in survivors of pediatric cancer: A report from the Childhood Cancer Survivor Study. Int J Radiat Oncol Biol Phys 2013;86:649-655. 4. Gajjar A, Chintagumpala M, Ashley D, et al. Risk-adapted craniospinal radiotherapy followed by high-dose chemotherapy and stemcell rescue in children with newly diagnosed medulloblastoma (St Jude Medulloblastoma-96): Long-term results from a prospective, multicentre trial. Lancet Oncol 2006;7:813-820. 5. Ris MD, Packer R, Goldwein J, et al. Intellectual outcome after reduced-dose radiation therapy plus adjuvant chemotherapy for medulloblastoma: A Children’s Cancer Group study. J Clin Oncol 2001; 19:3470-3476. 6. Taylor MD, Northcott PA, Korshunov A, et al. Molecular subgroups of medulloblastoma: The current consensus. Acta Neuropathol 2012; 123:465-472. 7. Sidney Kimmel Comprehensive Cancer Center. Study assessing the feasibility of a surgery and chemotherapy-only in children with Wnt positive medulloblastoma. 2018. https://clinicaltrials.gov/ct2/show/ NCT02212574. 8. St. Jude Children’s Research Hospital, Genentech, National Cancer Institute. A clinical and molecular risk-directed therapy for newly diagnosed medulloblastoma. 2023. https://clinicaltrials.gov/ct2/ show/NCT01878617. 9. Cavalli FMG, Remke M, Rampasek L, et al. Intertumoral heterogeneity within medulloblastoma subgroups. Cancer Cell 2017;31:737754.e6. 10. Morrissy AS, Cavalli FMG, Remke M, et al. Spatial heterogeneity in medulloblastoma. Nat Genet 2017;49:780-788. 11. Pickles JC, Hawkins C, Pietsch T, et al. CNS embryonal tumours: WHO 2016 and beyond. Neuropathol Appl Neurobiol 2018;44:151162. 12. Lastowska M, Jurkiewicz E, Trubicka J, et al. Contrast enhancement pattern predicts poor survival for patients with non-WNT/SHH medulloblastoma tumours. J Neurooncol 2015;123:65-73. 13. Koeller KK, Rushing EJ. From the archives of the AFIP: Medulloblastoma: A comprehensive review with radiologic-pathologic correlation. Radiographics 2003;23:1613-1637. 14. Bluml S, Margol AS, Sposto R, et al. Molecular subgroups of medulloblastoma identification using noninvasive magnetic resonance spectroscopy. Neuro Oncol 2016;18:126-131. 15. Patay Z, DeSain LA, Hwang SN, et al. MR imaging characteristics of wingless-type-subgroup pediatric medulloblastoma. AJNR Am J Neuroradiol 2015;36:2386-2393. 16. Perreault S, Ramaswamy V, Achrol AS, et al. MRI surrogates for molecular subgroups of medulloblastoma. AJNR Am J Neuroradiol 2014;35:1263-1269. 17. McGarry SD, Hurrell SL, Kaczmarowski AL, et al. Magnetic resonance imaging-based radiomic profiles predict patient prognosis in newly diagnosed glioblastoma before therapy. Tomography 2016;2:223-228. 12 Hara et al. 18. Coroller TP, Grossmann P, Hou Y, et al. CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma. Radiother Oncol 2015;114:345-350. 19. Vallières M, Kay-Rivest E, Perrin LJ, et al. Radiomics strategies for risk assessment of tumour failure in head-and-neck cancer. Sci Rep 2017;7:10117. 20. Kickingereder P, Götz M, Muschelli J, et al. Large-scale radiomic profiling of recurrent glioblastoma identifies an imaging predictor for stratifying anti-angiogenic treatment response. Clin Cancer Res 2016;22:5765-5771. 21. Li Q, Bai H, Chen Y, et al. A fully-automatic multiparametric radiomics model: Towards reproducible and prognostic imaging signature for prediction of overall survival in glioblastoma multiforme. Sci Rep 2017;7:14331. 22. Lao J, Chen Y, Li ZC, et al. A deep learning-based radiomics model for prediction of survival in glioblastoma multiforme. Sci Rep 2017; 7:10353. 23. Robles HA, Smirniotopoulos JG, Figueroa RE. Understanding the radiology of intracranial primitive neuroectodermal tumors from a pathological perspective: A review. Semin Ultrasound CT MR 1992; 13:170-181. 24. Sirin S, de Jong MC, Galluzzi P, et al. MRI-based assessment of the pineal gland in a large population of children aged 0-5 years and comparison with pineoblastoma: Part II, the cystic gland. Neuroradiology 2016;58:713-721. 25. Galluzzi P, de Jong MC, Sirin S, et al. MRI-based assessment of the pineal gland in a large population of children aged 0-5 years and comparison with pineoblastoma: Part I, the solid gland. Neuroradiology 2016;58:705-712. 26. Koci TM, Chiang F, Mehringer CM, et al. Adult cerebellar medulloblastoma: Imaging features with emphasis on MR findings. AJNR Am J Neuroradiol 1993;14:929-939. 27. Tortori-Donati P, Fondelli MP, Rossi A, et al. Medulloblastoma in children: CT and MRI findings. Neuroradiology 1996;38:352-359. 28. Bühring U, Strayle-Batra M, Freudenstein D, et al. MRI features of primary, secondary and metastatic medulloblastoma. Eur Radiol 2002;12:1342-1348. 29. de Carvalho Neto A, Gasparetto EL, Ono SE, et al. Adult cerebellar medulloblastoma: CT and MRI findings in eight cases. Arq Neuropsiquiatr 2003;61:199-203. 30. Malheiros SM, Carrete H Jr., Stávale JN, et al. MRI of medulloblastoma in adults. Neuroradiology 2003;45:463-467. 31. Griffiths PD, Coley SC, Romanowski CA, et al. Contrast-enhanced fluid-attenuated inversion recovery imaging for leptomeningeal disease in children. AJNR Am J Neuroradiol 2003;24:719-723. 32. Chawla A, Emmanuel JV, Seow WT, et al. Paediatric PNET: Presurgical MRI features. Clin Radiol 2007;62:43-52. 33. Rodallec M, Colombat M, Krainik A, et al. Diffusion-weighted MR imaging and pathologic findings in adult cerebellar medulloblastoma. J Neuroradiol 2004;31:234-237. 34. Liu HQ, Yin X, Li Y, et al. MRI features in children with desmoplastic medulloblastoma. J Clin Neurosci 2012;19:281-285. 35. Porto L, Kieslich M, Bartels M, et al. Leptomeningeal metastases in pediatrics: Magnetic resonance image manifestations and correlation with cerebral spinal fluid cytology. Pediatr Int 2010;52:541-546. 36. Freilich RJ, Krol G, DeAngelis LM. Neuroimaging and cerebrospinal fluid cytology in the diagnosis of leptomeningeal metastasis. Ann Neurol 1995;38:51-57. 37. Keil VC, Warmuth-Metz M, Reh C, et al. Imaging biomarkers for adult medulloblastomas: Genetic entities may be identified by their MR imaging radiophenotype. AJNR Am J Neuroradiol 2017;38: 1892-1898. 38. Zhao F, Li C, Zhou Q, et al. Distinctive localization and MRI features correlate of molecular subgroups in adult medulloblastoma. J Neurooncol 2017;135:353-360. 39. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: Extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-446. International Journal of Radiation Oncology Biology Physics 40. Sachdeva J, Kumar V, Gupta I, et al. Segmentation, feature extraction, and multiclass brain tumor classification. J Digit Imaging 2013; 26:1141-1150. 41. Orphanidou-Vlachou E, Vlachos N, Davies NP, et al. Texture analysis of T1- and T2-weighted MR images and use of probabilistic neural network to discriminate posterior fossa tumours in children. NMR Biomed 2014;27:632-639. 42. Rodriguez Gutierrez D, Awwad A, Meijer L, et al. Metrics and textural features of MRI diffusion to improve classification of pediatric posterior fossa tumors. AJNR Am J Neuroradiol 2014;35:1009-1015. 43. Fetit AE, Novak J, Peet AC, et al. Three-dimensional textural features of conventional MRI improve diagnostic classification of childhood brain tumours. NMR Biomed 2015;28:1174-1184. 44. Raleigh DR, Tomlin B, Buono BD, et al. Survival after chemotherapy and stem cell transplant followed by delayed craniospinal irradiation is comparable to upfront craniospinal irradiation in pediatric embryonal brain tumor patients. J Neurooncol 2017;131: 359-368. 45. Chang CH, Housepian EM, Herbert C Jr. An operative staging system and a megavoltage radiotherapeutic technic for cerebellar medulloblastomas. Radiology 1969;93:1351-1359. 46. Thompson EM, Hielscher T, Bouffet E, et al. Prognostic value of medulloblastoma extent of resection after accounting for molecular subgroup: A retrospective integrated clinical and molecular analysis. Lancet Oncol 2016;17:484-495. 47. Gajjar A, Sanford RA, Bhargava R, et al. Medulloblastoma with brain stem involvement: The impact of gross total resection on outcome. Pediatr Neurosurg 1996;25:182-187. 48. Louis DN, Perry A, Reifenberger G, et al. The 2016 World Health Organization Classification of Tumors of the Central Nervous System: A summary. Acta Neuropathol 2016;131:803-820. 49. Raleigh DR, Solomon DA, Lloyd SA, et al. Histopathologic review of pineal parenchymal tumors identifies novel morphologic subtypes and prognostic factors for outcome. Neuro Oncol 2017;19:7888. 50. Vallières M, Freeman CR, Skamene SR, et al. A radiomics model from joint FDG-PET and MRI texture features for the prediction of lung metastases in soft-tissue sarcomas of the extremities. Phys Med Biol 2015;60:5471-5496. 51. Zhou H, Vallières M, Bai HX, et al. MRI features predict survival and molecular markers in diffuse lower-grade gliomas. Neuro Oncol 2017;19:862-870. 52. Kickingereder P, Neuberger U, Bonekamp D, et al. Radiomic subtyping improves disease stratification beyond key molecular, clinical, and standard imaging characteristics in patients with glioblastoma. Neuro Oncol 2018;20:848-857. 53. Faraggi D, Reiser B. Estimation of the area under the ROC curve. Stat Med 2002;21:3093-3106. 54. Zweig MH, Campbell G. Receiver-operating characteristic (ROC) plots: A fundamental evaluation tool in clinical medicine. Clin Chem 1993;39:561-577. 55. Lasko TA, Bhagwat JG, Zou KH, et al. The use of receiver operating characteristic curves in biomedical informatics. J Biomed Inform 2005;38:404-415. 56. Obuchowski NA, Lieber ML. Confidence intervals for the receiver operating characteristic area in studies with small samples. Acad Radiol 1998;5:561-571. 57. Beier D, Kocakaya S, Hau P, et al. The neuroradiological spectra of adult and pediatric medulloblastoma differ: Results from a literaturebased meta-analysis. Clin Neuroradiol 2018;28:99-107. 58. Remke M, Hielscher T, Northcott PA, et al. Adult medulloblastoma comprises three major molecular variants. J Clin Oncol 2011;29: 2717-2723. 59. Pierce TT, Provenzale JM. Evaluation of apparent diffusion coefficient thresholds for diagnosis of medulloblastoma using diffusionweighted imaging. Neuroradiol J 2014;27:63-74. 60. Newton HB, Jolesz FA. Handbook of Neuro-oncology Neuroimaging. Cambridge, MA: Academic Press; 2008. Volume - Number - 2018 61. Kirks DR, Griscom NT. Practical Pediatric Imaging: Diagnostic Radiology of Infants and Children. Philadelphia, PA: Lippincott-Raven; 1998. 62. Koob M, Girard N, Ghattas B, et al. The diagnostic accuracy of multiparametric MRI to determine pediatric brain tumor grades and types. J Neurooncol 2016;127:345-353. 63. Aboian M, Kline C, Li Y, et al. Medu-40. Detection of nonenhancing recurrent medulloblastoma on diffusion weighted imaging. Neuro Oncol 2017;19(Suppl. 4). iv46-iv46. 64. Tixier F, Le Rest CC, Hatt M, et al. Intratumor heterogeneity characterized by textural features on baseline 18F-FDG PET images predicts response to concomitant radiochemotherapy in esophageal cancer. J Nucl Med 2011;52:369-378. 65. Smith AB, Rushing EJ, Smirniotopoulos JG. From the archives of the AFIP: Lesions of the pineal region: Radiologic-pathologic correlation. Radiographics 2010;30:2001-2200. 66. Fouladi M, Gajjar A, Boyett JM, et al. Comparison of CSF cytology and spinal magnetic resonance imaging in the detection of Pediatric embryonal brain tumor radiomics 67. 68. 69. 70. 71. 13 leptomeningeal disease in pediatric medulloblastoma or primitive neuroectodermal tumor. J Clin Oncol 1999;17:3234-3237. Gerber NU, Mynarek M, von Hoff K, et al. Recent developments and current concepts in medulloblastoma. Cancer Treat Rev 2014;40: 356-365. Harrison SK, Ditchfield MR, Waters K. Correlation of MRI and CSF cytology in the diagnosis of medulloblastoma spinal metastases. Pediatr Radiol 1998;28:571-574. Bennett J, Ashmawy R, Ramaswamy V, et al. The clinical significance of equivocal findings on spinal MRI in children with medulloblastoma. Pediatr Blood Cancer 2017;64:e26472. Morana G, Alves CA, Tortora D, et al. Added value of diffusion weighted imaging in pediatric central nervous system embryonal tumors surveillance. Oncotarget 2017;8:60401-60413. Fetit AE, Novak J, Rodriguez D, et al. Radiomics in paediatric neurooncology: A multicentre study on MRI texture analysis. NMR Biomed 2018;31:e3781.