Combining radiomics and deep convolutional neural network features from preoperative MRI for predicting clinically relevant genetic biomarkers in glioblastoma.

Background: Glioblastoma is the most common primary brain malignancy, yet treatment options are limited, and prognosis remains guarded. Individualized tumor genetic assessment has become important for accurate prognosis and for guiding emerging targeted therapies. However, challenges remain for widespread tumor genetic testing due to costs and the need for tissue sampling. The aim of this study is to evaluate a novel artificial intelligence method for predicting clinically relevant genetic biomarkers from preoperative brain MRI in patients with glioblastoma.
Methods: We retrospectively analyzed preoperative MRI data from 400 patients with glioblastoma, IDH-wildtype or WHO grade 4 astrocytoma, IDH mutant who underwent resection and genetic testing. Nine genetic biomarkers were assessed: hotspot mutations of IDH1 or TERT promoter, pathogenic mutations of TP53, PTEN, ATRX, or CDKN2A/B, MGMT promoter methylation, EGFR amplification, and combined aneuploidy of chromosomes 7 and 10. Models were developed to predict biomarker status from MRI data using radiomics features, convolutional neural network (CNN) features, and a combination of both.
Results: Combined model performance was good for IDH1 and TERT promoter hotspot mutations, pathogenic mutations of ATRX and CDKN2A/B, and combined aneuploidy of chromosomes 7 and 10, with receiver operating characteristic area under the curve (ROC AUC) >0.85 and was fair for all other tested biomarkers with ROC AUC >0.7. Combined model performance was statistically superior to individual radiomics and CNN feature models for prediction chromosome 7 and 10 aneuploidy, MGMT promoter methylation, and PTEN mutation. Conclusions: Combining radiomics and CNN features from preoperative MRI yields improved noninvasive genetic biomarker prediction performance in patients with WHO grade 4 diffuse astrocytic gliomas.

We evaluated artificial intelligence models for predicting glioblastoma genetics from MRI.

We assessed 9 different genetic biomarkers that are relevant for treatment or prognosis.

Combining radiomics and convolutional image features yielded improved performance.

Accurate radiogenomic assessment of preoperative brain MRI has the potential to improve care of patients with glioblastoma and could be particularly useful for helping to guide the use of emerging targeted therapies. This study demonstrates accurate preoperative radiogenomic assessment of several common glioblastoma genetic biomarkers including some that are relevant to investigational therapies. We further demonstrate that combining information from radiomics and deep learning feature extraction methods can lead to improved radiogenomic prediction accuracy compared to using either type of feature alone.

Glioblastoma is the most common primary brain malignancy but remains difficult to treat despite advances in surgical, radiation, and medical therapies.[1,2] The recent discovery that histologically defined glioblastoma encompasses several distinct tumor entities has accelerated progress in precision therapy targeting specific molecular biomarkers.[3,4] However, challenges remain for widespread tumor genetic testing due to costs and the need for adequate tissue sampling. In addition, in rare cases where biopsy is not possible or declined by the patient, glioblastoma is treated presumptively based on imaging alone, which precludes genetic testing. As our understanding of glioblastoma genetics improves, there is an increasing need for reliable, noninvasive methods for predicting genetic biomarkers relevant to prognosis and guiding targeted therapies.

Several prior studies have demonstrated that various clinically relevant glioblastoma genetic biomarkers can be inferred based on preoperative MRI using image analysis techniques. Most prior work has focused on image features extracted using radiomics or convolutional neural networks (CNNs) as the primary image-based predictors.[5-9] Radiomics features are a set of human-defined morphologic, grayscale, and pattern-based features that can be applied across multiple different imaging modalities and have proven useful for many different medical image-based inference tasks.[10] Radiomics features are relatively straightforward to define, conceptualize, and interpret, and are both standardized and reproducible.[11,12] In contrast, CNN-based features are abstract and often difficult to interpret multiscale features that are “learned” automatically using CNNs. CNN features are unique to each input dataset, which allows considerable versatility but also introduces susceptibility to overfitting and lack of reproducibility.[13] Relatively few prior studies have evaluated a combination of radiomics and CNN features for oncologic inference tasks[14-16] and studies specifically assessing genetic biomarker prediction in glioblastoma are lacking.

In this study, we present a hybrid approach for preoperative MRI-based glioblastoma genetic biomarker prediction, which combines both radiomics features and CNN features into a single unified artificial intelligence architecture. Our method leverages the strengths of both predefined radiomics features and learned CNN features to provide improved prediction performance. Predictive models were evaluated using a relatively large (n = 400) patient cohort with a diverse set of genetic biomarkers, many of which are useful either for determining prognosis or for guiding investigational targeted therapies.

Materials and Methods

Patient Population

All studies were performed in accordance with applicable guidelines and regulations and were approved by an institutional review board with a waiver for informed consent. The retrospective study cohort consisted of 400 adult patients with CNS WHO grade 4 diffuse astrocytic gliomas that had a final integrated diagnosis of either “Glioblastoma, IDH-wildtype” or “Astrocytoma, IDH-mutant” using the 2021 World Health Organization (WHO) Classification of Central Nervous System Tumors. All patients underwent preoperative MRI, initial tumor resection, and tumor genetic testing at the University of California San Francisco medical center between 2015 and 2021. The study cohort was identified by searching radiology and pathology databases (Nuance mPower) for all patients who received a preoperative planning MRI and had subsequent surgical pathology report within the following month with a final integrated diagnosis of glioblastoma, IDH-wildtype or astrocytoma, IDH-mutant, CNS WHO grade 4. Exclusion criteria included any prior history of brain tumor resection (n = 18) and incomplete or technically inadequate imaging studies (n = 28).

Genetic Biomarker Testing

Nine different pathogenic glioblastoma genetic biomarkers were evaluated: mutations or deletions of IDH1, TP53, PTEN, ATRX, TERT, and CDKN2A/B, MGMT promoter methylation, EGFR copy number amplification (including accompanying EGFRVIII rearrangement), and combined aneuploidy of chromosomes 7 and 10 (most commonly as the combination of trisomy 7 and monosomy 10).[17] Genetic biomarkers were chosen based on prevalence in the study cohort as well as their utility for determining tumor type (eg, IDH1, ATRX), informing chemotherapeutic management (eg, MGMT), determining prognosis (eg, IDH, PTEN), and/or guiding targeted therapies (eg, CDKN2A/B, EGFR). Gold standard molecular characterization was performed as part of routine clinical care at the time of diagnosis without knowledge of the imaging predictors developed during this study. Methods varied during the study period depending on the test availability and insurance reimbursement. Three hundred and five cases in the cohort were characterized with a capture-based targeted next-generation DNA sequencing panel as previously described.[18] For the remaining cases, molecular characterization was accomplished by fluorescence in situ hybridization (PTEN, EGFR) and immunohistochemical staining (IDH1, ATRX, TP53).[19,20] All IDH mutations identified by immunohistochemistry were confirmed by genetic sequencing. MGMT promoter methylation status was determined using a quantitative methylation sensitive PCR assay.[21]

Image Acquisition

A standardized preoperative brain tumor MRI protocol was performed at 3.0 Tesla (Discovery 750, GE Healthcare) with a dedicated 8-channel head coil (Invivo). The imaging protocol included 6 3D sequences: T2-weighted, T2-weighted FLAIR, susceptibility-weighted (SWI), pre- and postcontrast T1-weighted images, and arterial spin labeling (ASL) perfusion images, as well as 2 2D sequences: isotropic diffusion-weighted images (DWIs) and 55-direction high angular resolution diffusion imaging (HARDI). Complete parameters for the standardized MRI protocol are provided elsewhere.[22] Some image data from this study are publicly available via the 2021 BraTS challenge, and we plan to share the remainder as a separate collection.[23,24]

Image Preprocessing

All image data were deidentified and converted to NIfTI format using dcm2niix v1.0 (https://github.com/rordenlab/dcm2niix) using default parameters. HARDI data were eddy current corrected and processed using the Eddy and DTIFIT modules from the Functional MRI of the Brain Software Library (FSL) version 6.0.2 yielding mean diffusivity (MD), axial diffusivity (AD), radial diffusivity (RD), and fractional anisotropy (FA) images.[25] All images were registered and resampled to the 3D space defined by the T2/FLAIR image (1-mm isotropic resolution) using automated nonlinear registration (Advanced Normalization Tools 2.3.5).[26,27] Resampled coregistered data were then skull stripped (including face removal) using a previously described and publicly available method (https://github.com/ecalabr/brain_mask).[22] Finally, all images underwent N4 coil bias correction and intensity normalization.[28] Processing time varied from approximately 18 min to approximately 35 min per study using a standard desktop computer with a 16-core processor, 128 GB of RAM, and an Nvidia RTX Titan graphics processing unit.

Deep Learning-Based Automated Tumor Subcompartment Segmentation

A previously validated deep learning algorithm was used to generate automated 3D segmentations of 3 key components of glioblastoma that are seen on MRI: enhancing and central nonenhancing/necrotic tumor (together comprising the tumor core) and the surrounding T2/FLAIR abnormality typically referred to as “edema,” which may also contain infiltrative nonenhancing tumor outside of the tumor core.[29,30] All segmentations were manually corrected by a group of trained radiologists with varying experience (including authors EC and JR) using ITK-SNAP v3.8.0[31] and subsequently approved by 2 attending neuroradiologists with >15 years of experience each as part of the 2021 BraTS challenge.[24] Manual annotators were not provided any molecular characterization data.

Radiomics Feature Extraction

Radiomics features were extracted using PyRadiomics 2.2.0 command line tools with default extraction parameters.[32] Input images included 11 image contrasts: T1 pre, T1 post, T2, T2/FLAIR, SWI, DWI, ASL, MD, AD, RD, and FA. Input segmentations included 5 different tumor components: whole tumor, tumor core, and each of the 3 individual tumor compartments. The default set of shape features (n = 26), first-order grayscale features (n = 19), and higher-order grayscale features (n = 75) were extracted yielding 5300 radiomics features per patient.

Radiomics Feature Selection

An automated radiomics feature selection process was employed to select the 32 best features for inference using scikit-learn 0.24 with a 5-fold cross-validation approach to reduce selection bias. First, univariate feature selection with mutual information was used to select the 1024 best correlated features. Next, the best features for each cross-validation fold were ranked using recursive feature elimination with a random forest classifier, and the 32 features with the best average rank across folds were selected as the final feature set. Parameters for recursive feature elimination included 1000 estimators for the random forest classifier and a step size of 16 features for each feature elimination step.

Network Architecture

A CNN classifier (CNN limb) was constructed using TensorFlow 2.4. The CNN limb consisted of a 3D multiscale deep convolutional autoencoder with 1, 1, 2, and 2 bottleneck residual blocks per level, respectively, and a max-pool down-sampling layer (pool size = [2, 2, 2]) between each level.[33] Bottleneck blocks included leaky ReLU activation, batch normalization, and 40% feature dropout. The number of features per layer was set at 32 for the top level and doubled after each pooling step. A parallel random forest classifier limb was constructed for radiomics feature inference (radiomics limb) using scikit-learn 0.24. The random forest classifier parameters included 32 input features and 1000 estimators. Each limb yielded a single output logit, which was converted to a probability using the sigmoid logistic function. The 2 output probabilities (1 from the CNN limb and 1 from the radiomics limb) were averaged to create a final combined model probability. A probability threshold of P ≥ .5 was used to determine a positive result.

Network Inputs and Training

Network training was implemented with 5-fold cross-validation and a train/test split of 80%/20%. Separate networks were trained for each genetic biomarker. CNN limb inputs consisted of 96 × 96 × 96 voxel cubes encompassing the whole tumor (including surrounding FLAIR abnormality) after scaling to fit with a 5-voxel margin in each dimension. Data augmentation included random dimension swaps and rotations of ±0–90 degrees in each of the cardinal axes. Training data were sampled such that each batch consisted of an equal number of positive and negative cases. Model weights were initialized using the Glorot initializer.[34] Training comprised 40 epochs with a batch size of 10 cases on an Nvidia DGX-2 system. Neither training validation nor early stopping were used. Binary cross entropy loss between predictions and genetic biomarker results was minimized using the Adam optimizer.[35] Radiomics limb inputs consisted of the 32 best radiomics features identified during feature selection. Radiomics data were balanced using SMOTE feature synthesis.[36] Training was accomplished using the built in “fit” method from the sci-kit learn RandomForestClassifier class.

Network Performance Evaluation

Genetic biomarker prediction was evaluated using receiver operating characteristic (ROC) analysis as well as several different scalar metrics for the ROC best point: sensitivity, specificity, accuracy, precision, the F1 statistic, the Matthews correlation coefficient (MCC), and Youden’s index. Statistically significant differences between ROC curves were assessed using DeLong’s method with a threshold of P < .05.[37,38]

Results

Genetic biomarker testing was not uniform across the patient cohort. Gene test frequency, included variants, and prevalence in the study cohort are presented in Table 1.[17,39,40] In addition to the listed biomarkers, all cases were also tested for IDH2 p.R172 mutations, however, none were identified in the study cohort. Of note, 29 of 400 patients in the cohort had tumors with IDH1 p.R132 hotspot mutations corresponding with an integrated diagnosis of “Astrocytoma, IDH-mutant, CNS WHO grade 4”. While these tumors are now classified independently of IDH-wildtype glioblastomas in the 2021 WHO Classification of CNS Tumors, they were included in the study cohort to allow prediction of IDH mutations.

Table 1.

Genetic biomarker testing details for the study cohort

Genetic biomarker	Positive/tested	Male/ female	Age ± Std.	Variants included
IDH1	29/400	241/159	60 ± 13	Any p.R132 hotspot mutation (most often p.R132H)
MGMT	209/381	233/148	60 ± 13	Methylation of ≥2/17 promoter CpG sites
TP53	174/398	239/159	60 ± 13	Any pathogenic mutation or deletion
PTEN	210/399	240/159	60 ± 13	Any pathogenic mutation or deletion
ATRX	34/396	238/158	60 ± 13	Any pathogenic mutation or deletion
TERT	239/305	178/127	60 ± 14	Either c.-146C>T or c.-124C>T promoter hotspot mutation
CDKN2A/B	200/305	178/127	60 ± 14	Homozygous/biallelic deletion of CDKN2A and CDKN2B
EGFR	101/399	240/159	60 ± 13	Focal gene amplification (including accompanying EGFRVIII structural rearrangement)
Chromosomes 7 and 10	221/305	178/127	60 ± 14	Combined polysomy of 7 and monosomy of 10

For each biomarker, the number of positive cases and total number of tested cases are listed along with patient sex and average age. Specific genetic variants that were considered are listed in the far-right column.

Image Preprocessing, Segmentation, and Radiomics Feature Extraction

Figure 1 shows a schematic flow chart of the relevant image preprocessing, tumor segmentation, and radiomics feature extraction steps including all instances where user input was required. All study data were successfully processed through all steps of the pipeline. Data processing time varied depending on the amount of manual input required, but automated steps were accomplished in approximately 30 min per case.

Figure 1.

Flow chart of image processing steps. MR images (INPUT) were passed through several sequential automated image processing steps (rounded boxes) with minimal human input (diamond boxes). Each limb of the prediction model (convolutional neural network and radiomics) yields a single output logit, which is transformed into a probability using the sigmoid function. Final predictions (OUTPUT) were generated by averaging probabilities from the 2 limbs. Diffusion tensor derived contrasts include: mean diffusivity (MD), axial diffusivity (AD), radial diffusivity (RD), and fractional anisotropy (FA).

Deep Neural Network Architecture

A graphical representation of the deep CNN architecture used for genetic biomarker inference is provided in Supplementary Figure 1. The total number of trainable parameters was 3 564 279 for each CNN instance. Identical models were trained for each of the evaluated genetic biomarkers. Code for the CNN limb is available at https://www.github.com/ecalabr/gliomarad.

Model Inputs and Outputs

Example image inputs and genetic biomarker prediction outputs for a single patient are shown in Supplementary Figure 2. Image inputs included 3D image volumes of the entire tumor region for each of the 11 MRI contrasts included in the study. Outputs consist of probabilities of each individual genetic biomarker being present from the radiomics and CNN models, respectively, as well as a combined probability.

Genetic Biomarker Prediction Performance

ROC curves for each genetic biomarker are shown in Figures 2–4, and a summary of evaluation metrics is provided in Table 2. Both radiomics and CNN methods were able to predict each genetic biomarker with performance that could not be explained by chance alone. For some genetic biomarkers, such as CDKN2A/B homozygous deletion, the CNN feature-based model showed better performance by all metrics, while for others, such as TERT promoter mutation, the radiomics feature-based model showed better performance. Combined model receiver operating characteristic area under the curve (ROC AUC) was highest for ATRX (0.97) and IDH1 (0.96) mutations; however, these biomarkers had the lowest prevalence in the study cohort and therefore may be better assessed using metrics that account for class imbalance such as precision (0.42 and 0.65, respectively) or MCC (0.60 and 0.64, respectively). Combined model ROC AUC was lowest for MGMT promoter methylation (0.77) and PTEN mutation/deletion (0.77).

Figure 2.

Fivefold cross-validation receiver operating characteristic curves for radiomics, convolutional neural network, and combine models for predicting pathogenic alterations in ATRX, IDH1, and chromosomes 7 and 10. Best points are indicated with an outlined star. Average area under the curve is included in the legend of each subfigure. P values for DeLong’s test for ROC curve difference between the combined model and the radiomics (prad) and CNN (pcnn) models, respectively, are reported with “*” denoting statistical significance. CNN, convolutional neural network; ROC, receiver operating characteristic.

Figure 4.

Fivefold cross-validation receiver operating characteristic curves for radiomics, convolutional neural network, and combine models for predicting pathogenic alterations in TP53, MGMT, and PTEN. Best points are indicated with an outlined star. Average area under the curve is included in the legend of each subfigure. P values for DeLong’s test for ROC curve difference between the combined model and the radiomics (prad) and CNN (pcnn) models, respectively, are reported with “*” denoting statistical significance. CNN, convolutional neural network; ROC, receiver operating characteristic.

Table 2.

Performance metrics for the, radiomics, convolutional neural network (CNN), and combined models for each genetic biomarker

Biomarker	Model	AUC	p combined	Sens.	Spec.	Acc.	Prec.	F1	MCC	Youden
ATRX	Radiomics	0.93	.083	0.80	0.95	0.94	0.59	0.68	0.65	0.75
	CNN	0.95	.001*	1.00	0.83	0.84	0.34	0.51	0.53	0.83
	Combined	0.97		0.97	0.88	0.89	0.42	0.59	0.60	0.85
IDH1	Radiomics	0.89	.108	0.83	0.93	0.92	0.47	0.60	0.59	0.76
	CNN	0.96	.107	1.00	0.83	0.84	0.30	0.47	0.50	0.83
	Combined	0.96		0.86	0.94	0.93	0.52	0.65	0.64	0.80
Trisomy 7 Monosomy 10	Radiomics	0.79	.004*	0.69	0.76	0.71	0.88	0.77	0.40	0.45
	CNN	0.79	<.001*	0.71	0.75	0.72	0.88	0.78	0.41	0.45
	Combined	0.86		0.74	0.87	0.77	0.94	0.82	0.55	0.60
CDKN2A/B	Radiomics	0.76	<.001*	0.70	0.71	0.70	0.82	0.76	0.39	0.41
	CNN	0.85	.476	0.74	0.79	0.76	0.87	0.80	0.51	0.53
	Combined	0.86		0.82	0.73	0.79	0.85	0.84	0.54	0.55
TERT	Radiomics	0.87	.504	0.84	0.75	0.82	0.92	0.88	0.54	0.59
	CNN	0.75	<.001*	0.71	0.68	0.70	0.89	0.79	0.33	0.38
	Combined	0.85		0.87	0.67	0.83	0.90	0.89	0.52	0.54
EGFR amplification	Radiomics	0.77	.124	0.70	0.74	0.72	0.59	0.64	0.42	0.44
	CNN	0.72	<.001*	0.68	0.66	0.66	0.51	0.58	0.32	0.33
	Combined	0.80		0.77	0.71	0.73	0.58	0.66	0.45	0.47
TP53	Radiomics	0.75	.115	0.68	0.73	0.71	0.50	0.58	0.37	0.40
	CNN	0.71	<.001*	0.54	0.78	0.71	0.50	0.52	0.31	0.32
	Combined	0.79		0.75	0.68	0.70	0.48	0.59	0.39	0.42
MGMT promoter methylation	Radiomics	0.70	<.001*	0.63	0.70	0.66	0.77	0.69	0.32	0.33
	CNN	0.73	.005*	0.72	0.63	0.68	0.75	0.73	0.34	0.34
	Combined	0.77		0.70	0.74	0.72	0.81	0.75	0.43	0.44
TEN	Radiomics	0.74	.014*	0.73	0.69	0.71	0.72	0.73	0.42	0.42
	CNN	0.69	<.001*	0.68	0.62	0.65	0.66	0.67	0.30	0.30
	Combined	0.77		0.75	0.69	0.72	0.72	0.74	0.44	0.44

Acc., accuracy; F1, F1 statistic; MCC, Matthew’s correlation coefficient; Prec., precision; ROC AUC, receiver operating characteristic area under the curve; Sens., sensitivity; Spec., specificity; pcombined: P value for DeLong’s test for ROC curve difference compared to the combined model. “*” denotes P < .05.

Fivefold cross-validation receiver operating characteristic curves for radiomics, convolutional neural network, and combine models for predicting pathogenic alterations in CDKN2A/B, TERT promoter, and EGFR. Best points are indicated with an outlined star. Average area under the curve is included in the legend of each subfigure. P values for DeLong’s test for ROC curve difference between the combined model and the radiomics (prad) and CNN (pcnn) models, respectively, are reported with “*” denoting statistical significance. CNN, convolutional neural network; ROC, receiver operating characteristic.

Comparison With Isolated Radiomics or CNN Feature-Based Models

The combined radiomics and CNN feature model showed improved ROC AUC compared to either the radiomics or CNN models individually for all evaluated genetic biomarkers. Combined model ROC AUC was statistically superior (DeLong’s P < .05) to at least one of the individual models (radiomics or CNN) for all genetic biomarkers except for IDH1 mutations and was superior to both individual models (radiomics and CNN) for combined chromosome 7 and 10 aneuploidy, MGMT promoter methylation, and PTEN mutation/deletion. Combined model performance by other metrics was variable but generally better than or comparable to the better of the 2 individual models. In some cases, select combined model metrics were slightly worse than the better of the 2 individual models (eg, the precision metric for ATRX).

Discussion

In this study, we evaluated an artificial intelligence architecture that combines both radiomics and CNN image features into a single model for predicting genetic biomarkers in glioblastoma. We found that combining these 2 types of image features resulted in significantly improved genetic biomarker prediction performance in most cases and synergistic improvement in some cases. In our dataset, prediction performance of the individual radiomics and CNN models varied significantly between genetic biomarkers with the radiomics model performing better in some cases and the CNN model performing better in others. This highlights the fact that radiomics and CNN feature relevance may vary significantly depending on the specific inference task. One potential advantage of the proposed combined approach is that it can balance the contribution of radiomics and CNN features and can perform equally well for tasks that are better suited for either type of image feature.

In addition to presenting a new method for genetic biomarker prediction in glioblastoma, this study provides new insight into the feasibility of image-based genetic prediction for several genetic biomarkers that have not previously been well evaluated. For example, our model was able to predict CDKN2A/B homozygous deletion with a ROC AUC of 0.86 and accuracy of 79%. This genetic biomarker is present in approximately 60% of IDH-wildtype glioblastomas, and can be targeted using small molecule inhibitors of CDK4/6, which have shown promise for treating other tumor types and are currently being evaluated in glioblastoma.[41,42] These results suggest that image-based genetic inference may be useful for a variety of different glioblastoma genetic biomarkers beyond IDH1 and MGMT, which have largely dominated prior related work. However, image-based genetic inference can only be accurate if the target biomarker produces an imaging phenotype. It is conceivable, that some genetic alterations may not yield discernable imaging phenotypes and therefore may not be amenable to image-based inference. By similar logic, genetic alterations that frequently occur together may be difficult to distinguish. For example, in our study cohort, only 8 patients had either ATRX or IDH1 mutation, but not both, and the combined model was only able to identify 4/8 (50%) of these discordant mutations. The low prevalence of these genetic biomarkers also highlights the issue of class imbalance, which can lead to overly optimistic ROC AUC values. For example, the best ROC AUC values in our study were achieved for biomarkers with the lowest prevalence (ie, ATRX, IDH1). Model performance on biomarkers with low prevalence may therefore be better evaluated using metrics that account for class imbalance such as precision or MCC.

One major limitation of this study is the lack of external validation. Unfortunately, glioblastoma genetic testing is still not widely adopted, particularly for newer less studied genetic biomarkers. In addition, there is wide variation in preoperative brain tumor MR imaging protocols in terms of sequences, field strength, and acquisition dimensionality. As result, many publicly available glioblastoma MRI datasets do not provide genetic data and/or include only a limited number of MRI sequences (typically T1, T2, FLAIR, and postcontrast imaging). For these reasons, there are no publicly available datasets that could reasonably be used to externally validate the models proposed in this study. It should be noted that recent efforts to create a generalizable model for MGMT promoter methylation status prediction in a multi-institution glioblastoma MRI dataset yielded relatively poor performance, which is presumed to be related to data heterogeneity and/or the limited variety of available MRI sequences.[24] Given these findings, presented results should be interpreted as “best case scenario” as performance on external data would almost certainly be worse. However, as the clinical importance of glioblastoma genetics continues to grow and preoperative MRI protocols continue to improve, it seems likely that data relevant to this study will become more widely available in the future. To this end, the codebase for this study has been made publicly available at https://www.github.com/ecalabr/gliomarad and image data are publicly available via the 2021 BraTS challenge.[23,24]

As the clinical potential of glioblastoma radiogenomics continues to mature, there is increasing need for new predictive modeling methods that can better utilize available image data. Our results suggest that there may be a role for combining radiomics and CNN features for improved genetic biomarker prediction accuracy in glioblastoma. This approach has shown promise for other radiologic/oncologic inference tasks and may ultimately prove to be an important step toward integration of image-based genetic inference into clinical workflows. Future work will be necessary to evaluate the optimal architectures for combining radiomics and CNN features and for assessing generalizability and portability to other datasets.

Click here for additional data file.