Prediction of Molecular Mutations in Diffuse Low-Grade Gliomas using MR Imaging Features

Diffuse low-grade gliomas (LGG) have been reclassified based on molecular mutations, which require invasive tumor tissue sampling. Tissue sampling by biopsy may be limited by sampling error, whereas non-invasive imaging can evaluate the entirety of a tumor. This study presents a non-invasive analysis of low-grade gliomas using imaging features based on the updated classification. We introduce molecular (MGMT methylation, IDH mutation, 1p/19q co-deletion, ATRX mutation, and TERT mutations) prediction methods of low-grade gliomas with imaging. Imaging features are extracted from magnetic resonance imaging data and include texture features, fractal and multi-resolution fractal texture features, and volumetric features. Training models include nested leave-one-out cross-validation to select features, train the model, and estimate model performance. The prediction models of MGMT methylation, IDH mutations, 1p/19q co-deletion, ATRX mutation, and TERT mutations achieve a test performance AUC of 0.83 ± 0.04, 0.84 ± 0.03, 0.80 ± 0.04, 0.70 ± 0.09, and 0.82 ± 0.04, respectively. Furthermore, our analysis shows that the fractal features have a significant effect on the predictive performance of MGMT methylation IDH mutations, 1p/19q co-deletion, and ATRX mutations. The performance of our prediction methods indicates the potential of correlating computed imaging features with LGG molecular mutations types and identifies candidates that may be considered potential predictive biomarkers of LGG molecular classification.

Molecular alterations (IDH mutation, 1p/19q co-deletion, ATRX, and TERT mutation), grade (II and III), and clinical data are downloaded from the Genomic Data Commons Data Portal (https://portal.gdc.cancer.gov/). Clinical data are de-identified by the Health Insurance Portability and Accountability Act of 1996 (HIPAA). The distribution of the data is as follows: (i) IDH mutation: 85 Mutant (of which 27 cases are co-deleted) and 23 wild-type (WT), (ii) 1p/19q co-deletion: 27 co-deletion and 81 non-co-deletion, (iii) ATRX status: 43  Methodology. In this study, we introduce different molecular prediction models based on fractal and multi-resolution fractal texture features and other MR imaging features. These molecular models include the IDH, 1p/19q co-deletion, MGMT, ATRX, and TERT prediction. A classical way to avoid overfitting is to divide the dataset into training, validation and testing datasets 42 . The dataset is randomly partitioned into n pairs (partitions) of training (75% of the entire dataset = 81 cases) and testing (25% of the entire dataset = 27 cases). A balanced distribution of the target molecular mutation is ensured in the training and testing sets in each molecular prediction model. The features are extracted from multimodality MRI sequences of the tumor volume in the training partition. Then, a recursive feature selection is performed to select the number of features and validated with Leave-One-Out Cross-Validation (LOOCV). The selected features are then trained using an Extreme Gradient Boosting (XGBoost) method along with LOOCV. Then, a prediction performance is evaluated using the testing partition.
Furthermore, we study the efficacy of fractal and multi-resolution fractal texture features (e.g., piecewise-triangular prism surface area (PTPSA), multi-resolution fractional Brownian motion (mBm), and Holder Exponent (HE)) extracted from tumor volumes on the performance of the molecular mutation prediction models. Figure 1 shows the overall pipeline of the proposed LGG-XGBoost prediction model for different molecular mutations in LGG.
The computational algorithm of the PTPSA, mBm, and the HE is found in 36,44,46 . Furthermore, six histogram-based statistics (mean, variance, skewness, kurtosis, energy, and entropy) features are also extracted from the different tumor sub-regions (edema, enhancing tumor, and necrosis), respectively.
In addition, we extracted 12 volumetric features: the volume of the whole tumor, the volume of the whole tumor with respect to the brain, the volume of sub-regions (edema, enhancing tumor, and necrosis) divided by the whole tumor, the volume of sub-regions (edema, enhancing tumor, and necrosis) divided by the brain, the volumes of the enhancing tumor and necrosis divided by the edema, the summation of the volume of the edema and enhancing tumor, the volume of the edema divided by the summation of the volume of enhancing tumor and necrosis, and the volume of the necrosis divided by the summation of the volume of the edema and enhancing tumor. Finally, nine-volume properties (area, bounding box, centroid, perimeter, major axis length, minor axis length, eccentricity, orientation, solidity, and extent) are extracted from the tumor volume and from three viewpoints (x, y, and z-axes) of the tumor and its sub-regions (edema, enhancing tumor, and necrosis).
Texture features are analyzed using MATLAB-based software developed by Vallières et al. 43 . Fractal characterization, multi-resolution fractal characterization, HE characterization, and volumetric features are analyzed using MATLAB-based in-house software. Molecular mutation prediction model and feature selection. The molecular mutation prediction model is performed on the training set using nested LOOCV to avoid an optimistic performance estimate. Recursive feature selection is performed in the inner loop to find optimum features set. In each mutation prediction analysis, feature selection using RFS is done using the training set that is (75%) of the whole 108 cases. Because we repeat the prediction analysis of training and testing n times, eventually each case of the 108 cases would have appeared in the training sets. The LOO cross-validated performance of the molecular mutation prediction model is estimated in the outer loop. The molecular mutation prediction model is performed using the R statistical packages Caret and XGBoost (www.r-project.org).
Recursive feature selection (RFS) is implemented by first fitting a Random Forest (RF) model to all features. Each feature is ranked by its importance, and the least important features are removed from the current feature set. Then, this step is repeated recursively until the optimum features set that has the best performance is reached. In our implementation of recursive feature selection, the number of features in the features' sets are 2, 3, 5, 7, 9, 11, 13, and 15 features. In addition, the best performance is determined by maximizing the area under the receiver operating characteristic (ROC) metric. The features' set that provides the combination of features that maximize the area under the ROC (AUC) is chosen for training in the prediction model. Using recursive feature selection, the maximum number of selected features is fifteen, so that the training samples (eighty-one cases) are at least 5 times the number of features to reduce model overfitting.
In our study, XGBoost is utilized as a classification and prediction model using the optimum features set as input and molecular mutation information as the target output. XGBoost 47 is an advanced tree boosting Figure 1. The general outline of the LGG-XGBoost molecular prediction model. Note that this analysis is repeated n times (randomly chosen) to generate a more reliable performance.
supervised machine learning technique that is effective in handling imbalanced datasets. XGBoost is widely used in classification and regression tasks. For a given dataset D with s samples and m features , a tree ensemble model uses K additive functions to predict the output as follows, where x i is the feature/input vector, y i is the target/output variable, and f x ( ) is a function in the functional space F, and F is a set of all possible classification and regression trees. One of the major advantages of using XGBoost is that XGBoost provides L1 and L2 regularization. L1 regularization handles sparsity, whereas L2 regularization reduces overfitting. In addition, we choose XGBoost because it is known for handling an imbalanced dataset. A detailed mathematical derivation of the XGBoost algorithm is found in Chen et al. 47 .
The final molecular mutation prediction model (that is used for the testing set) is obtained by fitting the optimum features' set that maximize the performance in the inner loop (over all the outer cross-validation loops). Note, if there are more than one feature sets maximizes the inner loop performance, then the common feature between the features' sets are used. The prediction performance of the final molecular mutation model is tested using the paired testing sets (partitions).
Finally, in order to study the efficacy of fractal and multi-resolution fractal texture features used in this study (e.g., PTPSA, mBm, and Holder Exponent) on the performance of the proposed prediction models as shown in Fig. 1, we perform molecular prediction analyses with and without these texture features, respectively. The whole process in Fig. 1 is repeated n times independently with n different training/testing set pairs. The n number of repetitions is a random number between 10 and 15 that are generated for each model.
Evaluation. The molecular models are validated using separate testing sets and the prediction performance (test performance) of the trained XGBoost model is estimated using AUC, sensitivity, and specificity. After n times of independent repetitions, ANOVA test is used to compare the difference in the prediction performance between two models with and without the fractal and multi-resolution fractal texture in the prediction models. In addition, ANOVA is used to analyze the significant association between features and the different molecular mutations. The survival groups that are formed using the significant features are compared using Kaplan-Meier curves and the log-rank test. The hazard ratio of features is determined using the Cox proportional hazards model and assessed using the likelihood-ratio test. Finally, the evaluation step for Survival is conducted using R statistical packages.

Results
Around 680 imaging features are extracted from multimodality MRI sequences. Recursive feature selection is used to find the optimum number of significant features for each molecular mutation prediction model. Our analysis of the different prediction models are repeated independently n times with different training and testing pairs (partitions). Table 1 displays the number of repetitions n, LOOCV performance, and the test performance of the different prediction models when including/removing texture characterization of the fractal and multi-resolution fractal of PTPSA, mBm, and Holder Exponent characterization.
MGMT methylation model. The most frequent features that are selected are the necrosis width in the Z planar, the histogram entropy of the mBm characterization on the whole tumor, and size ratio between the enhancing tumor and the necrosis. The necrosis width in the Z planar and the histogram entropy of the mBm characterization on the whole tumor features are significantly (ANOVA test, p-value < 0.05) associated with methylated MGMT, whereas the size ratio between enhancing tumor and necrosis is associated significantly with un-methylated MGMT as shown in Fig. 2A. www.nature.com/scientificreports www.nature.com/scientificreports/ The LOOCV performance and the prediction performance on the testing set for predicting the MGMT methylation status using imaging features are illustrated in Table 1 and Fig. 2B. The prediction performance using the test partitions achieves an AUC, a sensitivity, and a specificity of 0. 0.83 ± 0.04, 0.93 ± 0.05, and 0.73 ± 0.13, respectively. Removal of our fractal and multi-resolution fractal features from MGMT methylation prediction model have an effect on the prediction performance on the testing set. AUC and specificity drop significantly (ANOVA test, p-value = 0.003, and 0.01, respectively) when the fractal features are removed (Fig. 2B).
IDH mutation model. Our analysis reveals that the tumor correlation, the vertical orientation of edema major axis (the angle between the edema major axis and the vertical axis), size ratio between the enhancing tumor and the necrosis, and the complexity of holder exponent of the tumor are among the most frequently selected features to predict IDH-mutated status in LGG. Higher values of the complexity of the holder exponent of the tumor, the size ratio between the enhancing tumor and the necrosis, and higher values of the vertical orientation of edema major axis associate significantly (ANOVA test, p-value < 0.005) with WT IDH status. Whereas the tumor correlation associates significantly (ANOVA test, p-value < 0.005) with mutated IDH status as illustrated in Fig. 3A. Figure 3B shows the clustering of IDH status using the most frequent features in Fig. 3A. The clustering between the mutated IDH and WT IDH is demonstrated using t-Distributed Stochastic Neighbor Embedding 48 (tSNE).
In addition, the tumor correlation, the vertical orientation of edema major axis, and the complexity of holder exponent of the tumor features carry a hazard ratio [HR] of 0.562 (95% CI, 0.381-0.828), 2.655 (95% CI, 1.617-4.36), and 1.553 (95% CI, 1.165-2.07) with a likelihood ratio test p-value = 0.005, 0.0001, 0.008, respectively. Because these features are continuous features, the HRs interpolate as follows: the risk of death increases (or decreases if HR < 1) by (HR-1) × 100% for every 1-standard deviation increase in that feature. The LOOCV and the test performance of the proposed IDH models are illustrated in Table 1. The prediction performance using the testing partitions achieves an AUC, sensitivity, and specificity of 0.84 ± 0.03, 0.90 ± 0.06, and 0.79 ± 0.09, respectively. Note that the AUC and specificity of the IDH status prediction model drop significantly to 0.75 ± 0.07 and 0.66 ± 0.18 (ANOVA test, p = 0.0001 and p = 0.028, respectively) after removing features extracted from fractal and the multi-resolution modeling (Fig. 3C).
1p/19q co-deletion model. The necrosis upper-left bounding box location, the histogram entropy of mBm characterization of the whole tumor, and the horizontal coordinate of necrosis centroid are the most frequent optimum features that are selected in our proposed 1p/19q codeletion models. These three features show significance (ANOVA test, p-value < 0.0001) associated with the existence of the 1p/19q co-deletion. Our analysis shows frontal tumors are associated significantly (ANOVA test, p-value < 0.0001) with 1p/19q co-deletion mutations (Fig. 4A). The performance of the proposed co-deletion prediction LOOCV model is illustrated in Table 1. The 1p/19q co-deletion performance using the test partitions achieve an AUC of 0.80 ± 0.04, a sensitivity of 0.75 ± 0.08, and a specificity of 0.85 ± 0.06. In addition, the efficacy of our fractal and multi-resolution fractal texture features on the performance of the co-deletion prediction model is significant as shown in (Fig. 4B and Table 1). The AUC and the sensitivity of the co-deletion prediction model drop significantly (ANOVA test, p-value of 0.024 and 0.029, respectively) after removing features extracted from our fractal and multi-resolution fractal features in the 1p/19q co-deletion prediction model to 0.75 ± 0.07 and 0.67 ± 0.12 (without fractal features). Figure Fig. 5A. Higher values of information of correlation are associated significantly (ANOVA test, p-value < 0.001) with ATRX wild-type. Whereas Higher values of histogram mean are associated significantly (ANOVA test, p-value < 0.001) with mutated ATRX. The ATRX prediction model achieves prediction performance of an AUC of 0.70 ± 0.09, a sensitivity of 0.70 ± 0.06, and a specificity of 0.83 ± 0.10 using the test partitions. Removing features extracted from our fractal and multi-resolution fractal modeling from the ATRX prediction model has a significance specificity drop to 0.68 ± 0.18 performance of the model with p-value = 0.03 (ANOVA test) as shown in Fig. 5B. TERT mutation model. A review of the most frequently selected imaging features of TERT mutation prediction model, we notice that the information content of correlation of the tumor volume, the edema upper-left bounding box location, and the inverse difference moment of HE characterization of tumor volume are the most frequently selected features (Fig. 5C). The inverse difference moment of HE characterization of tumor volume offers HR of 0.612 per standard deviation (95% CI, 0.405-0.924) with a likelihood ratio test p-value = 0.027. The TERT prediction models' performances are illustrated in Table 1. The TERT prediction performance using the test partitions achieves an AUC of 0.82 ± 0.04, a sensitivity of 0.77 ± 0.12, and a specificity of 0.86 ± 0.09, respectively. Removing our fractal and multi-resolution fractal texture modeling has no significant effect on the TERT performance of prediction models when using the test partitions as shown in Table 1.

IDH/co-deletion model. Molecular classification based on the status of IDH mutations and 1p/19q
co-deletion results in distinguishing three LGG molecular subtypes that have a distinct clinical behavior: IDH WT, IDH mutant with 1p/19q co-deleted, and IDH mutant with 1p/19q non-co-deleted. In this work, we combine these two mutations and perform a 3-class prediction (i.e., IDH/co-deletion prediction model) using the methodology illustrated previously and shown in Fig. 1.
The IDH mutation model and the 1p/19q co-deletion model show their superior performance when using fractal and multi-resolution fractal modeling features along with other non-fractal (i.e., texture and volumetric) features. Consequently, we develop 3-class IDH/co-deletion models utilizing fractal and other non-fractal features. The whole process of training and testing is repeated 5 times. The distribution of the three molecular  www.nature.com/scientificreports www.nature.com/scientificreports/ subtypes is as follows: 23 cases represent IDH WT, 27 cases represent IDH mutant with 1p/19q co-deleted, and 58 cases represent IDH mutant with 1p/19q non-co-deleted.
Our analysis using the Chi-square test confirms a significant association (p-value = 0.004) between IDH status and 1p/19q codeletion. Additionally, our analysis shows that IDH WT patients carry HR of 3.1 (likelihood ratio test p-value = 0.007) and have significantly shorter survival when compared to IDH mutated patients (19.9 vs. 65.7 months, log-rank test p-value = 0.004). The association between IDH status and overall survival remains significant after stratifying for 1p/19q codeletion (likelihood ratio test p-value = 0.005). Table 2 shows the performance of the 3-class training and testing of the IDH/co-deletion model where the sensitivity and the specificity are reported per class, multiclass AUC is calculated as described in 49 , and the overall accuracy is calculated based on the number of correctly classified in all classes to the total number of cases.
Note this simultaneous prediction of multiple genes may be more clinically valuable when compared to conventional single gene prediction models.

Discussion
The 2016 WHO classification of diffuse LGGs heavily weighs molecular mutations classifying primary brain tumors with particular importance assigned to IDH mutation, 1p/19q co-deletion, ATRX mutation, TERT mutations, and MGMT methylation. Our study on diffuse LGG is largely able to predict the presence of these important molecular mutations based on MR imaging features. Therefore, prediction of tumor aggressiveness (based on molecular mutations) may be achieved through non-invasive imaging features as an adjunct to traditional visual morphologic diagnosis and invasive tissue sampling.
In this work, the number of originally extracted imaging features (six hundred eighty features) is higher than the number of samples (eighty-one cases) in the training dataset, which may cause overfitting. To address possible   www.nature.com/scientificreports www.nature.com/scientificreports/ overfitting, we implement feature selection in the training model that offers a maximum of fifteen selected features. Figure 6 illustrates the effect of the number of features on the cross-validated performance of the different mutation prediction models. The average AUC performances and the standard error of the different prediction models improve when the number of features is greater than 9. Note that the standard error mostly plateaus when the number of features varies between 9-15 (standard error reflects instability).
For the fractal and multi-resolution fractal texture model in Table 1, the AUC predictive performance of MGMT, 1p/19q co-deletion, and ATRX models drop to 0.83 ± 0.04, 0.80 ± 0.04, and 0.70 ± 0.09, respectively. This statistically non-significant drop in performance (ANOVA test, p = 0.076, p = 0.073, and p = 0.071 respectively) when compared to their AUC cross-validated performance, may suggest minimal overfitting for these models. The AUC predictive performances of IDH and TERT models are almost comparable to their cross-validated performance as shown in Table 1 that suggests there is no overfitting in these two models.
For the non-fractal models in Table 1, the AUC predictive performance of IDH, 1p/19q co-deletion, and TERT models drop to 0.75 ± 0.07, 0.75 ± 0.07, and 0.78 ± 0.07 respectively. This statistically non-significant drop (ANOVA test, p = 0.182, p = 0.056, and p = 0.062 respectively) when compared to their AUC cross-validated performance, may suggest minimal overfitting for these models. However, the poor predictive performances of the non-fractal MGMT and non-fractal ATRX models when compared with their non-fractal cross-validated performances are a sign of overfitting.
Note "optimally chosen" features in the non-fractal models (in Figs. 2-4) are not selected by simply replacing the fractal features by alternate features (or by using the same number of predictors as in the fractal models). In each mutation model (fractal or non-fractal), RFS (and thus selecting "optimally chosen" features) is performed independently. In our implementation of RFS for fractal or non-fractal models in Figs. 2-4, the possible number of features in the features' sets may be 2, 3,5,7,9,11,13, and 15, respectively. The maximum number of selected features is set to 15 such that the training samples (81 cases) are at least 5 times the number of features to reduce the possibility of model overfitting.
Finally, we compare the performances of our prediction models with a list of state-of-the-art studies as illustrated in Table 3. However, a direct comparison between the performances of our proposed models and these studies may not be relevant because of the different datasets.
Our study on MGMT methylation prediction shows that MGMT methylation correlates with high values of fractal texture features such as histogram entropy of mBm for tumor volume. Entropy measures randomness or uncertainty in the tumor. The analysis reveals that high histogram entropy of mBm associates with the less aggressive methylated MGMT status and carries HR of 0.579 per standard deviation (95%CI, 0.345, 0.969) with a likelihood ratio test p-value = 0.035. The study further shows that the size ratio between enhancing tumor and necrosis correlates significantly with un-methylated MGMT, which indicates that the high aggressive MGMT un-methylated LGG, the higher the values of the size ratio. The MGMT methylation prediction study by Kanas www.nature.com/scientificreports www.nature.com/scientificreports/ et al. 50 for patients with GBM reports the size of the tumor with respect to necrosis as one of the significant features. Our analysis shows that removing the texture features such as fractal and the multi-resolution fractal (of PTPSA, mBm, and Holder Exponent) characterization is significant on the AUC and specificity performance of the MGMT methylation model. The GBM study conducted by Kanas et al. 50 (Table 3) proposes an MGMT prediction model using volumetric, morphological, and locational MR imaging features, respectively. In our study, we use texture features and volumetric features. Moreover, the whole process of the prediction model in our current study including feature extraction is automated, unlike the proposed work by Kanas et al. 50 . Another study by Han et al. 51 (Table 3), the authors use a bi-directional convolutional recurrent neural network to predict MGMT methylation status. A major difference between our MGMT prediction model and the method proposed by Han et al. 51 is that our model mainly utilizes quantitative imaging features that can be correlated with tumor biology.
In addition, our IDH mutation prediction model indicates that the tumor correlation associates significantly with mutated IDH and offers HR of 0.562 per standard deviation with a likelihood ratio test p-value = 0.005. In addition, our analysis shows that the complexity of HE of enhancing tumor associates significantly with WT IDH status with HR of 1.553 per standard deviation with a likelihood ratio test p-value = 0.008. Complexity is related to the visual information content and the shape of the object. Texture with higher information content and with a large number of edges are considered complex 52 . This outcome is in agreement with another gliomas study by Wang et al. 53 which reports that the enhancement patterns predict the prognosis in IDH1 mutations in Anaplastic gliomas. Our analysis also shows that the size ratio between enhancing tumor and necrosis is a significant predictor feature of the IDH status. This feature is also a significant predictor of MGMT status, which can be explained by the fact that MGMT methylation is associated with IDH status as reported by Mukasa et al. 54 . In the IDH prediction model by Yu et al. 55 (Table 3), the authors use 110 imaging features and SVM to classify IDH status with Grade II glioma patients. Even though the dataset we use in the IDH mutation prediction is not the same as the dataset is used by Yu et al. 55 , the dataset used in our study is more diverse with data from both Grade II and III, and this reflects higher reliability of the performance of the IDH status prediction model. A different study by Ding et al. 56 (Table 3) on 76 LGG patients utilizes MR imaging features along with MR spectroscopic data to predict IDH mutations using a binary logistic regression model. The authors achieved the best performance when utilizing MR spectroscopic data. When comparing the performance of our IDH prediction and the best performance of Ding et al. 56 , our model achieves better AUC, sensitivity, and specificity as illustrated in Table 3.
Furthermore, our study on the 1p/19q co-deletion prediction model indicates that the location of the upper-left necrosis bounding box and horizontal coordinate of the necrosis centroid (illustrated in Fig. 4C) are among the most predictive features. This outcome is in agreement with different studies [57][58][59] which report that gliomas with 1p/19q co-deletion are associated with the tumor location. In addition, our analysis shows that higher values of histogram entropy of mBm texture of tumor volume are associated significantly with the existence of 1p/19q co-deletion. Moreover, our analysis reveals that removing the texture representation of fractal and multi-resolution fractal from the 1p/19q co-deletion prediction model decreases the AUC and specificity  Table 3. Comparison between our proposed molecular mutations models and state-of-the art glioma grading models.
significantly. The test prediction performance of the 1p/19q co-deletion prediction model drops (after removing the fractal features). A study by Akkus et al. 60 (Table 3) with LGG patients (where each patient has 3 MRI slices) proposes 1p/19q co-deletion prediction using a convolutional neural network (CNN). Their method achieves better sensitivity; however, our method achieves slightly better specificity. In addition, Akkus et al. 60 do not consider the global information of the tumor, since their dataset uses only 3 slices of the MRI sequence of each patient as input, not the whole volume of the tumor. Another recent study by van der Voort et al. 61 (Table 3) utilizes MR imaging features along with patients' age and sex using an SVM classifier to predict 1p/19q co-deletion in LGG patients. The authors use 284 LGG patients for training and another 129 LGG patients for testing. Their analysis reveals that the cranial/caudal location of the tumor is one of the most important features in predicting 1p/19q co-deletion. Comparing the performance of our 1p/19q co-deletion prediction and the performance of van der Voort et al. 61 , our 1p/19q co-deletion prediction model outperforms their model as illustrated in Table 3.
Our analysis of ATRX-status prediction shows that tumor information measure of correlation imaging feature and histogram mean tumor volume are the most frequently selected features. Higher values of information measure of correlation are associated significantly with WT ATRX status. This is in agreement with an ATRX mutation prediction study by Li et al. 62 (Table 3) in patients with low-grade glioma, where the authors use MRI texture feature and LASSO regression model. In their model, the information measure of correlation is one of the features that is used to predict ATRX mutation. In addition, our analysis shows that the tumor information measure of correlation is one of the most frequent features in the TERT model as well. This can be explained by the fact that ATRX and TERT mutations are mutually exclusive.
The TERT prediction analysis shows that tumor information-measure of correlation and upper-left edema bounding box are the most frequently selected features. Higher values of these two features are significantly associated with mutated TERT status. The information-measure of correlation assesses the correlation/dependency between two gray-levels using mutual information content. High values of Information measure of correlation are associated with mutated TERT. In addition, our analysis suggests that the higher values of Inverse difference moment of HE associates significantly with WT TERT and offers HR = 0.612 per standard deviation with a likelihood ratio test, p-value = 0.03. Inverse difference moment measures local homogeneity. High values of inverse difference moment of HE tumor predict the less aggressive WT TERT. Recently, Wang et al. 63 (Table 3) explore survival prediction and TERT mutations in 39 LGG (30 WT, and 9 Mutant) patients and achieve a LOOCV AUC of 0.874 (95% CI: 0.756-0.992). The authors propose a TERT prediction model using 24 imaging features selected using Principle Component Analysis (PCA) and classified using the Partial Least Squares (PLS). While their method achieves a slightly higher AUC, the sensitivity, specificity, or the confusion matrix are not provided for such an imbalanced dataset.
Overall, our analysis shows that the necrosis location and the necrosis volume-related features are very important (most frequently selected features) in MGMT, IDH, and 1p/19q co-deletion prediction. Edema volume-related features are very important in IDH and TERT prediction models. Fractal features have a significant effect on MGMT, IDH, 1p/19q co-deletion, and ATRX prediction models. Further analysis on the most frequent features in each prediction model, we notice that the effect of thresholding the value of standardized feature around the median can stratify the 108 cases significantly (log-rank test, p-value < 0.05) into two survival groups ( Fig. 7A-C). The features and the median survival of each group are: • the size ratio between the enhancing tumor and necrosis stratifies the 108 cases into two groups with a median survival of 87.4 months vs 30.7 months, • the correlation of the tumor volume stratifies the 108 cases into two groups with a median survival of 114 months vs 46 months, • and the vertical orientation of edema major axis stratifies the 108 cases into two groups with a median survival of 114 vs 44 months.
In summary, this study presents molecular prediction model designs from traditional MRI data based on the 2016 update of the WHO classification of LGG of the CNS. Our prediction model performance shows promise when compared to different methods and models in the literature. An association among computed MR imaging features and the molecular mutations LGG was established. The methods discussed in our study are important www.nature.com/scientificreports www.nature.com/scientificreports/ steps towards non-invasive imaging classification of diffuse LGG based on molecular mutations prior to invasive tissue sampling. In this work for the first time in literature, we hypothesize that fractal and multi-resolution fractal features have an association with molecular prediction. The feature selection using RFS and the subsequent prediction results in Table 1 confirm our hypothesis by showing the efficacy of these fractal features in glioma prediction. Therefore, this work may be considered as a validation of previously hypothesized fractal biomarkers, and, hence, may have potential for generalizability for other types of tumors.