Abstract
BACKGROUND AND PURPOSE: Molecular biomarker identification increasingly influences the treatment planning of pediatric low-grade neuroepithelial tumors (PLGNTs). We aimed to develop and validate a radiomics-based ADC signature predictive of the molecular status of PLGNTs.
MATERIALS AND METHODS: In this retrospective bi-institutional study, we searched the PACS for baseline brain MRIs from children with PLGNTs. Semiautomated tumor segmentation on ADC maps was performed using the semiautomated level tracing effect tool with 3D Slicer. Clinical variables, including age, sex, and tumor location, were collected from chart review. The molecular status of tumors was derived from biopsy. Multiclass random forests were used to predict the molecular status and fine-tuned using a grid search on the validation sets. Models were evaluated using independent and unseen test sets based on the combined data, and the area under the receiver operating characteristic curve (AUC) was calculated for the prediction of 3 classes: KIAA1549-BRAF fusion, BRAF V600E mutation, and non-BRAF cohorts. Experiments were repeated 100 times using different random data splits and model initializations to ensure reproducible results.
RESULTS: Two hundred ninety-nine children from the first institution and 23 children from the second institution were included (53.6% male; mean, age 8.01 years; 51.8% supratentorial; 52.2% with KIAA1549-BRAF fusion). For the 3-class prediction using radiomics features only, the average test AUC was 0.74 (95% CI, 0.73–0.75), and using clinical features only, the average test AUC was 0.67 (95% CI, 0.66–0.68). The combination of both radiomics and clinical features improved the AUC to 0.77 (95% CI, 0.75–0.77). The diagnostic performance of the per-class test AUC was higher in identifying KIAA1549-BRAF fusion tumors among the other subgroups (AUC = 0.81 for the combined radiomics and clinical features versus 0.75 and 0.74 for BRAF V600E mutation and non-BRAF, respectively).
CONCLUSIONS: ADC values of tumor segmentations have differentiative signals that can be used for training machine learning classifiers for molecular biomarker identification of PLGNTs. ADC-based pretherapeutic differentiation of the BRAF status of PLGNTs has the potential to avoid invasive tumor biopsy and enable earlier initiation of targeted therapy.
ABBREVIATIONS:
- AUC
- area under the receiver operating characteristic curve
- FGFR
- fibroblast growth factors receptor
- GG
- ganglioglioma
- GLMD
- gray-level dependence matrix
- ML
- machine learning
- NF1
- neurofibromatosis 1
- NPV
- negative predictive value
- OvR
- one versus the rest
- PA
- pilocytic astrocytoma
- pLGG
- pediatric low-grade glioma
- PLGNT
- pediatric low-grade neuroepithelial tumor
- RF
- random forests
- WHO
- World Health Organization
Pediatric low-grade neuroepithelial tumors (PLGNTs) are the most common pediatric CNS tumors, representing approximately 40% of all pediatric brain tumors.1,2 According to the 5th Edition of the World Health Organization (WHO) classification, PLGNTs are grade 1 and 2 neoplasms described in 3 groups: diffuse pediatric low-grade gliomas (pLGGs), circumscribed astrocytic gliomas, and glioneuronal and neuronal tumors.3 The genetic and molecular features have been incorporated in this updated classification and, along with the histologic phenotype, relate to different prognoses and new targeted therapies.4,5
Most commonly, PLGNTs present with genetic changes causing upregulation of the Ras/mitogen-activated protein kinase (RAS/MAPK) pathway, usually fusions or mutations in the B-Raf proto-oncogene, serine/threonine kinase (BRAF) gene. The chromosomal alteration in BRAF fusion involves the duplication of the BRAF oncogene, followed by its insertion into one of several fusion targets, most often the K1AA1549 gene (KIAA1549-BRAF fusion).6 BRAF V600E point mutations activate BRAF, causing deregulation in the mitogen-activated p.V600E protein kinase pathway.7 Other common mutations include neurofibromatosis 1 (NF1), fibroblast growth factors receptors 1 and 2 (FGFR 1/2),2,8 and MYB/MYBL1 rearrangements.9⇓-11 The most common histology for the KIAA1549-BRAF fusion and BRAF V600E mutation tumors is pilocytic astrocytoma (PA) and ganglioglioma (GG), respectively.2,12 However, except for molecularly defined tumors, molecular subtypes are often found across various entities.2,13
Even though surgery is the preferred treatment for most PLGNTs when possible, more than one-half of PLGNTs cannot be completely resected, requiring subsequent pharmacologic therapy, often leading to high morbidity.5,14 Genetic differences are essential for therapeutic decision-making, with BRAF and MEK inhibitors gaining popularity and new trials being developed.15⇓-17 Notably, BRAF inhibitors exhibit efficacy in BRAF V600E mutation tumors, while KIAA1549-BRAF fusion tumors might exhibit progression due to pathway paradoxic activation when treated with these inhibitors. Instead, KIAA1549-BRAF fusion tumors respond positively to MEK inhibitors.15 However, implementing these new therapeutic strategies requires tissue sampling and molecular testing, mostly restricted to academic reference centers. Therefore, neuroimaging surrogates would be valuable tools for prognosis and treatment-planning.2
Radiomics has been applied to determine tumor histology, prognostication, and molecular classification (radiogenomics) in pediatric neuro-oncology.18 Zhang et al19 developed an MR imaging–based machine learning (ML) decision path that identified the 4 molecular subgroups of medulloblastomas. Wagner et al17 demonstrated that FLAIR sequences predicted molecular subgroups of pLGG with a high area under the receiver operating characteristic curve (AUC) using a supervised ML approach. In addition, qualitative ADC features have recently been shown to predict BRAF V600E mutations among pLGGs.16
Despite the increased application of ML in pediatric neuroradiology, to date, there are no studies assessing radiomic features of ADC images in PLGNTs. Given the widespread use of ADC images and their relatively straightforward interpretation in the context of neoplastic cellularity and prognostication,16 we hypothesized that KIAA1549-BRAF fusion, BRAF V600E mutation, and other non-BRAF PLGNTs can be differentiated quantitatively using the ADC map segmentations and that radiomics can improve diagnostic accuracy. Specifically, we aimed to develop and validate a radiomics ADC signature predictive of the molecular status of PLGNTs in pediatric patients using a bi-institutional cohort.
MATERIALS AND METHODS
Patients
This retrospective study was approved by the institutional review boards or research ethics boards of the 2 participating academic institutions: The Hospital for Sick Children, Toronto, Canada and Lucile Packard Children's Hospital, Stanford, Palo Alto, California. Due to the retrospective nature of the study, informed consent was waived by the local institutional review board/research ethics board. An interinstitutional data-transfer agreement was obtained for data-sharing. Patients were identified from the respective electronic health record database from January 2009 to January 2018. Patient inclusion criteria were the following: 1) 0–18 years of age, 2) availability of molecular information on BRAF status in histopathologically confirmed PLGNTs, and 3) availability of pretherapeutic brain MR imaging with non-motion-degraded FLAIR and DWI sequences, including ADC maps. Patients with spinal cord tumors were excluded.
Molecular Analysis
We used a targeted testing protocol. BRAF fusion status was determined using an nCounter Metabolic Pathways Panel (NanoString Technologies) or fluorescence in situ hybridization. BRAF p.V600E mutation was determined using immunohistochemistry, Droplet Digital (Bio-Rad Laboratories) polymerase chain reaction, as previously described.14,20 For 233 patients, molecular analysis was performed with formalin-fixed paraffin-embedded tissue obtained during surgery. Nineteen patients had molecular subtyping based on frozen tissue. Fluorescence in situ hybridization was further used to detect MYB, MYBL1, and FGFR2 fused transcripts. Droplet Digital polymerase chain reaction was used to detect FGFR1 alteration, CDKN2A deletion, and IDH1 mutation. DNA sequencing was used for samples negative for our targeted testing protocol.
MR Imaging Acquisition, Data Retrieval, and Image Segmentation
Patients from the first academic institution (The Hospital for Sick Children, Toronto, Ontario, Canada) underwent brain MR imaging at 1.5T or 3T across various vendors (Signa, GE Healthcare; Achieva, Philips Healthcare; Magnetom Skyra, Siemens). Sequences acquired included axial DWI with ADC map calculation (b = 0 and 1000 s/mm2, TR/TE, 6000–11,000/70–110 ms; 3- to 4-mm section thickness, 1 -to 3-mm gap), axial 2D T2 FLAIR, axial or coronal 2D T2-weighted FSE, and axial or sagittal 3D T1-weighted and axial or sagittal gadolinium-based contrast agent–enhanced T1-weighted sequences. Patients from the second academic institution (Lucile Packard Children’s Hospital, Palo Alto, California) underwent brain MR imaging at 1.5T or 3T from a single vendor (Signa or Discovery 750; GE Healthcare). MR imaging was performed using the brain tumor protocol of the institution, which included 2D axial T2-weighted FSE, 2D axial or sagittal precontrast T1-weighted spin-echo, 2D axial T2 FLAIR, DWI with ADC calculation (b =0 and 1000 s/mm2, TR/TE, 7000–10,000/80–110 ms, 3- to 5-mm section thickness, 1- to 3-mm gap), and 2D axial gadolinium-based contrast agent–enhanced T1-weighted spin-echo sequences.
All MR imaging data were extracted from the respective PACS and de-identified for further analysis. Tumor segmentation was performed by a pediatric neuroradiology fellow with 3 years of neuroradiology research experience using 3D Slicer (Version 5.0.3; http://www.slicer.org).21 Semiautomated tumor segmentation on ADC images was performed with the level tracing effect tool. Given that the tumors are better demonstrated on FLAIR images, these sequences were used as a guide to achieving the segmentation on the ADC images. The segmentation included both cystic and solid components of the tumor. This semiautomatic approach has been found superior to multiuser manual delineation concerning the reproducibility and robustness of results.22 The senior author, a board-certified neuroradiologist, confirmed the final and proper placement of VOIs. Figure 1 demonstrates the segmentation process.
Axial FLAIR (A and D), ADC maps (B and E), and manual tumor segmentation using semiautomatic tools in 3D Slicer (C and F). Upper row: a 15-year-old adolescent boy with a supratentorial BRAF-mutated low-grade glioma. Lower row: a 3-year-old boy with a cerebellar BRAF-fused low-grade glioma.
Radiomics Feature-Extraction Methodology
We used min-max normalization to scale the ADC images to [0, 1] and used the pyradiomics library with default hyperparameters to extract the radiomics features. A total of 688 radiomic features from the 3D VOIs in the ADC images were extracted for each patient. The radiomics features included shape, histogram, and texture features with and without wavelet-based filters. All features are listed in the Online Supplemental Data. Bias field correction was performed before segmentation to standardize the range of all image features.23,24 Preprocessing and radiomics feature extraction in 3D Slicer and other software have been described elsewhere.25⇓-27 Clinical variables, including age, sex, and location, were appended to the data set.
Machine Learning and Statistical Analysis
We used random forests (RF) as the classification algorithm.28 Previously, Wagner et al29 demonstrated that RF perform best with limited data sets compared with other ML algorithms, including XGBoost (https://xgboost.readthedocs.io/en/stable/), and neural network architectures. The hyperparameter grid space was defined according to the Online Supplemental Data. We conducted 100 experiments and used an AUC to evaluate the classifiers. We randomly split our data into test (20%, 50 patients) and development (dev, 80%, 202 patients) sets in each iteration in a loop with N repetitions. Subsequently, a filtration algorithm was trained on the dev set and applied to the test set to eliminate highly correlated features (correlation coefficient >0.95). To measure the validation performance, models with the best hyperparameter set were trained and validated 100 times (N_val) on random dev data splits with a 75/25 ratio. Finally, an instance of the model with the best hyperparameters was trained on the entire dev set and evaluated on the test cohort. The whole process was repeated 100 times (N); thus, we have 100 test AUCs. The final evaluation of the model was conducted on the unseen test cohort.
We trained the RF classifiers in the supervised learning framework with 3 different inputs: 1) radiomics alone, 2) clinical variables (age, sex [binary], tumor location [binary]), and 3) radiomics and clinical variables together. A one versus the rest (OvR) multiclass AUC was used as the evaluation metric. Additionally, we monitored average per-class AUCs. We also captured the most meaningful features of the classifiers and selected the top 10 features across the experiments.
“Large p, Small n” is a well-known problem in ML, in which the number of features has a higher order in comparison with the number of examples. Radiomics-based ML is often affected by Large p, Small-n and is prone to overfitting. A crucial step is allocating an unseen cohort of data for test (compared with conventional methods such as standard cross-validation) to tackle the issue. To that end, we had separate train, validation, and test cohorts. In terms of feature selection, we filtered low-variance features (feature variance <0.05) and highly correlated features (correlation coefficient >0.95). Last, we repeated the experiments using different model initializations and data splits to ensure that the test performance was reproducible and overfitting was indeed avoided.
The analysis was conducted using Python 3.9.7 and scikit-learn 1.3.0 (https://scikit-learn.org/stable/index.html). For feature extraction, we used PyRadiomics 3.0.1 (https://pyradiomics.readthedocs.io/en/v3.0.1/changes.html) with SimpleITK 2.1.1.2 (https://simpleitk.org/doxygen/v1_0/html/).
RESULTS
Patients
The data of 422 patients were screened for inclusion. The data sets of 168 patients were excluded due to the absence of DWI-derived ADC images (134 for not having ADC maps and 30 for having DTI). Four patients were excluded due to marked susceptibility artifacts from hemorrhage, and 2 were excluded for errors in radiomics processing (Fig 2). After all preprocessing and radiomics extraction steps, 252 patients were included (first institution, n = 229, second institution, n = 23), comprising 135 boys (53.6%); mean age, 8.01 (SD, 4.88) years. Patient demographics and pathologic information consisting of age, sex, histologic diagnosis, BRAF molecular status, and anatomic location of the tumor (supra-versus-infratentorial) are shown in Table 1.
Flow chart of the study.
Patient demographics
Among the non-BRAF tumors (n = 79, 31.2%), we identified the following molecular markers: FGFR1 (n = 10, 3.9%), NF1 (n = 5, 2%), MYB proto-oncogene like 1 (MYBL1) (n = 4, 1.6%), MYB proto-oncogene (MYB) (n = 4, 1.6%), FGFR2 (n = 2, 0.8%), isocitrate dehydrogenase 1 (IDH1) (n = 2, 0.8%), histone H3.3 K27M (n = 2, 0.8%), transforming acidic coiled-coil containing protein 1 (TACC1) (n = 2, 0.8%), platelet-derived growth factor receptor α (PDGFRA) (n = 2, 0.8%), PIK3CA (n = 2, 0.8%), QKI-RAF1 (n = 2, 0.8%), as well as a single case (0.4%) of anaplastic lymphoma kinase (ALK), neurotrophic tyrosine receptor kinase (NTRK), MYB-QKI rearrangement, mitogen-activated protein kinase 1 (MAPK1), chloride voltage-gated channel 6 (CLCN6), FYCO-RAF1, G protein subunit α 11, and GOPC-ROS1. In addition, cyclin-dependent kinase inhibitor 2A (CDKN2A) gene deletions were identified in 29 tumors (11.5%): 12 KIAA1549-BRAF fusions (4.76%), 7 BRAF V600E mutations (2.77%), and 10 non-BRAF tumors (3.96%).
Radiomics Model Evaluation
Given that one of our cohorts did not have 1 tumor subtype, the bi-institutional cohort was analyzed together, and 252 patients were randomly split into 80% development and 20% test cohorts in each experiment, as illustrated in the radiomics pipeline in Fig 3. We used 3-class classifiers and achieved the following results (OvR, AUC) over 100 experiments (different data splits).
The repetitive classification approach.
The top discriminative features are shown in Table 2. We encountered the first-order and gray-level dependence matrix (GLMD) classes among our top discriminative features. The mean, kurtosis, and 90th percentile in this class differ among the 3 subgroups. In addition, the dependence variance, large dependence emphasis, and large dependence high gray-level emphasis differed among our subgroups. The most important clinical feature was tumor location (supra- and infratentorial).
Top predictive radiomics and clinical features
Figure 4 shows the average AUC across all groups. By means of only the radiomics features, the average test AUC was 0.74 (95% CI, 0.73–0.75) using the OvR method. When the input was restricted to clinical features, the average test AUC dropped to 0.67 (95% CI, 0.66–0.68). The combination of both radiomics and clinical features improved the predictive performance to an AUC of 0.77 with (95% CI, 0.75–0.77).
Boxplots of test AUCs. Note that P values are calculated on the basis of Student t tests.
The diagnostic performance of the per-class test AUC in identifying KIAA1549-BRAF fusion, BRAF V600E mutation, and non-BRAF tumors among the other subgroups is shown in Table 3. It was higher in identifying KIAA1549-BRAF fusion tumors among the other subgroups (AUC = 0.81 for the combined radiomics and clinical features versus 0.75, 0.74 for BRAF V600E mutation, and non-BRAF, respectively).
Per-class and overall OvR test of AUC performance of the models
DISCUSSION
Using a bi-institutional cohort, we trained and tested radiomics models predictive of KIAA1549-BRAF fusion, BRAF V600E mutations, and non-BRAF alterations in PLGNTs. The optimal RF model achieved an accuracy AUC of 0.77 when combining radiomics and clinical features. In addition, we showed that the diagnostic performance was higher in differentiating KIAA1549-BRAF fusion from BRAF V600E mutation and non-BRAF tumors, regardless of which features were used (radiomics, clinical, and combined). To our knowledge, this is the first study that investigated ADC-based radiomics features in PLGNTs.
PLGNTs are a heterogeneous group of tumors arising predominantly from the glial cell lineage, including astrocytic, oligodendrocytic, or mixed neuronal and glial lineage.9 These neoplasms are classified as WHO grades 1 and 2 and are grouped into 3 categories in the fifth WHO classification: circumscribed astrocytic gliomas, glioneural and neuronal tumors, and pediatric-type diffuse low-grade gliomas (those with characteristic ill-defined infiltrative margins).3 PAs are the most common well-circumscribed PLGNT (up to 85%) and are by far the most common brain tumor type with KIAA1549-BRAF fusion. Therefore, the KIAA1549-BRAF fusion is a useful putative diagnostic marker, particularly for PAs, which can show the neuropathologic features of necrosis and microvascular proliferation, also seen in high-grade gliomas.30 The BRAF V600E mutation, on the other hand, is present in a more heterogeneous group of neoplasms with a worse prognosis, such as GG, dysembryoplastic neuroepithelial tumors, pleomorphic xanthoastrocytomas, astroblastomas, and diffuse low-grade gliomas MAPK pathway-altered.31 These 2 BRAF tumor subgroups also differ clinically because those with the V600E mutation tend to be supratentorial, while those with KIAA1549-BRAF fusion are usually located in the brainstem or cerebellum.9
DWI reflects the free movement of water molecules in tissue.32 Highly cellular tumors often show decreased ADC values.33 Most studies use ROI analysis to evaluate brain tumors,34 though it represents only a part of the lesion and is highly subjective. A better alternative is ADC histogram calculation, which analyzes the entire lesion and may show tissue heterogeneity;35 however, this technique is limited because different tissue components may have overlapping signal intensities.36 Alternatively, ML classifiers using radiomics features can extract data from larger volumes of images, increasing reproducibility, providing signatures in different CNS neoplasms, and assessing therapeutic interventions.36⇓⇓-39 ML models are associated with randomness; and especially with small data sets, it is crucial to measure the randomness of the models. Additionally, the reproducibility of radiomics studies is a major concern that demands a systematic evaluation. In this research, we followed the OpenRadiomics (https://arxiv.org/abs/2207.14776)38,40 protocol to achieve reproducible results.
Ramaglia et al41 first described the association between BRAF V600E and lower relative ADC values. Similar diagnostic performances were predicted by different parameters, such as relative ADC mean (AUC = 0.83, P < .001) and relative ADC minimum (AUC = 0.85, P < .001).41 Trasolini et al16 and Shrot et al35 also reported the association of BRAF V600E and lower ADC values in cohorts of 70 and 40 patients, respectively, suggesting higher tumor aggressiveness and cellularity in this subgroup.42 With a larger cohort, our results support the use of ADC in subgroup differentiation and add that radiomics features can predict the BRAF status. In addition, we exploit the entire set of radiomics features, including but not limited to histogram parameters.16,35,41 We also expand the use of ADC in subgroup differentiation by adding a third non-BRAF cohort to the prediction.
RF is an ensemble learning method that combines decision trees and handles various data types, including missing data.43 In our study, first-order and GLMD classes were among our top discriminative features. First-order statistics describe the distribution of voxel intensities within the image region through commonly used and basic metrics. The mean, kurtosis, and 90th percentile in this class differ among the 3 subgroups. Larger mean, 90th percentile, and kurtosis have been associated with reduced overall survival in glioblastomas.44 In addition, the mean, median, and 90th percentiles among pediatric patients were lower in the BRAF V600E subgroup.35 These metrics reflect the microstructure at the cellular level.45 In particular, the low diffusivity of BRAF V600E-mutant pLGG may be explained by the dense, compact architecture with fiber-rich regions.46 GLMD quantifies gray-level dependencies in an image, defined as the number of connected voxels in a certain distance dependent on a central voxel. The dependence variance, large dependence emphasis, and large dependence high gray-level emphasis in this class differed among our subgroups. These are the main features quantifying tumor heterogeneity, which might be missed through the human eyes on MR imaging.47 Texture features have been shown to differ among brain tumors48 and predict IDH status in low-grade gliomas.49 Also, the GLMD was the main feature class among the top discriminative features in KIAA1549-BRAF fusion and BRAF V600E mutation tumors,17 reflecting the molecular and histologic differences of these tumors.50
In our study, tumor location (supra- versus infratentorial) was the most important clinical discriminative feature across the experiments. Many studies support this finding,9,16,17,35 because KIAA1549-BRAF fusion is more often described in the posterior fossa and BRAF V600E mutation in the cerebral hemispheres and diencephalon.35 Sex and age were not significantly different among groups.16 Other features have also been reported, such as ill-defined margins35 and a higher number of cystic components.41 However, the significance of these findings varies in the literature. Ramaglia et al,41 for instance, did not find significant differences in tumor margins, the presence of hemorrhage/calcification, and contrast enhancement among BRAF subgroups in a cohort of 56 participants. Conversely, Shrot et al35 showed that KIAA1549-BRAF fusion tumors had significantly better-defined margins compared with BRAF V600E mutation tumors and that BRAF subgroups showed more cystic components than the NF1 subgroup in a cohort of 51 participants.
Wagner et al17 trained and validated the radiomics features of FLAIR sequences to predict BRAF status in 115 patients with pLGGs, achieving an AUC of 0.85 on the external validation data set. Similarly, Xu et al51 extracted features from FLAIR sequences in a cohort of 113 pediatric patients, building a model with a test set AUC of 0.87. We believe that the lower AUC for the ADC values in our study is because we expanded the cohort to include rare subtypes and compared each group versus the rest 3 times. In addition, FLAIR sequences show better margins and facilitate semiautomatic segmentation with good reproducibility.18 Finally, pLGGs are heterogeneous neoplasms, and FLAIR sequences may better depict this heterogeneity, including cystic and solid components, margins, and the range of signal changes.17
In this study, we noted that the prediction of the KIAA1549-BRAF fusion cohort achieved the highest AUC in the per-class analysis. This may be, in part, explained by the larger number of cases in this cohort. In addition, the heterogeneity of histopathologically defined entities was lower in this cohort compared with the other 2 cohorts. Most cases are PA. While it is conceivable to assume that the driver behind this prediction is histopathologic rather than molecular, we also noted that the predictive performance of our algorithm remained high with the other 2 cohorts in the per-class analysis. In contrast to the KIAA1549-BRAF fusion cohort, those cohorts were heterogeneous with respect to histopathology. Therefore, we speculate that radiomics can extract predictive information from ADC maps beyond histopathology.
Our study has limitations that need to be considered when interpreting the results. First, due to the retrospective and bi-institutional nature, there was heterogeneity in the DWI sequence acquisition, including using different imaging parameters, field strengths, and scanner vendors. The incorporation of heterogeneous training and testing data was previously shown to be an important component in the development of a robust and predictive ML model. In fact, training of heterogeneous imaging data might increase the generalizability of our findings.52 Second, our study segmented only ADC images. Incorporating additional sequences could further increase the model performance. Third, we opted to merge the 2 cohorts rather than keep them separated due to an imbalanced distribution of the third class (non-BRAF). Fourth, the larger number of cases of KIAA1549-BRAF fusion tumors may partially explain the higher AUC among the “fusion versus the rest” per-class analysis. Last, although our study cohort is one of the largest PLGNT cohorts to date, we did not analyze the predictive performance of radiomics features in the subgroup-specific determination of histopathologically defined tumors across molecular markers. Large international MR imaging data sets are necessary for robust subanalyses.
CONCLUSIONS
We explored the feasibility of radiomics features to predict BRAF status in PLGNT using ADC-based MR images in a bi-institutional cohort. The optimal RF model included a combination of clinical and radiomic features and achieved diagnostic accuracy. In particular, the highest accuracy was achieved to identify the KIAA1549-BRAF fusion tumor subgroup. Differentiating these molecular subgroups in PLGNT is paramount for newer target therapies.42 Future investigations with additional imaging sequences, such as T2WI and contrast-enhanced T1WI, may improve prediction accuracy.
Footnotes
C. Hawkins was supported by the Canadian Cancer Society (grant No. 702296) and the Canadian Institute of Health Research (grant No. 159805).
Disclosure forms provided by the authors are available with the full text and PDF of this article at www.ajnr.org.
References
- Received November 10, 2023.
- Accepted after revision January 16, 2024.
- © 2024 by American Journal of Neuroradiology
