Predicting Deletion of Chromosomal Arms 1p/19q in Low-Grade Gliomas from MR Images Using Machine Intelligence

Magnetic resonance (MR) imaging is a non-invasive medical imaging technique that provides outstanding soft tissue contrast and has become the standard imaging technique for brain tumor diagnosis [1]. Gliomas are the most frequent primary brain tumors originating in the brain [2]. World Health Organization (WHO) classifies them into four grades based on their aggressiveness. Low-grade gliomas (LGG), also named diffuse low-grade and intermediate-grade gliomas (WHO grades II and III), include oligodendrogliomas, astrocytomas, and oligoastrocytomas [3‐5]. Compared to high-grade gliomas (HGG: WHO grade IV, glioblastoma), LGG are less aggressive tumors with better prognosis. A subgroup of LGG will progress to glioblastoma (HGG, grade IV), but other subgroups will progress slower or remain stable [4, 6‐8]. In addition, some LGG are sensitive to therapy, and their survival ranges from 1 to 15 years [6, 8]. Presently, treatment includes observation, surgery, radiotherapy, and chemotherapy either separately or in combination [4]. Although histological grading of tumors is the gold standard for diagnosis and subsequent treatment planning, it is known that histopathological diagnosis lacks information about other tumor properties (e.g., genomic biomarkers) that can impact optimal therapy options [3]. Therefore, other tests of LGG (e.g., molecular biomarkers testing) are also obtained to improve treatment planning. Several studies [9‐12] have shown that codeletion of 1p/19q chromosome arms is a strong prognostic molecular marker for positive tumor response to chemotherapy and radiotherapy in LGG and associated with longer survival. Therefore, predicting 1p/19q status is crucial for effective treatment planning of LGG.

Currently, determining 1p/19q status requires surgical biopsy typically followed by fluorescence in-situ hybridization (FISH) [13] to identify chromosomal deletion. Several studies have shown that imaging can predict 1p19q status from MR images or positron emission tomography (PET) images. Fellah et al. [9] presented univariate analysis and multivariate random forest models to determine 1p/19q status from multimodal MR images including conventional MR images, diffusion-weighted imaging (DWI), perfusion-weighted imaging (PWI), and MR spectroscopy. DWI, PWI, and MRI spectroscopy showed no significant difference between tumors with and without 1p/19q loss in their study. They concluded that inclusion of DWI, PWI, and MR spectroscopy was not useful for determining 1p/19q status compared with conventional MR images. Jansen et al. [10] presented detection of 1p/19q status from [¹⁸F] fluoroethyltyrosine-PET (FET-PET) images. They derived several biomarkers from PET images and correlated with 1p/19q status, and showed that these biomarkers do not reliably predict the status of 1p/19q in individual patients. Iwadate et al. [11] studied detection of 1p/19q codeletion from ¹¹C–methionine PET images and concluded that ¹¹C–methionine PET might help discriminate tumors with and without 1p/19q codeletion preoperatively. Bourdillion et al. [12] presented prediction of anaplastic transformation in grade 2 oligodendrogliomas based on MR spectroscopy and 1p/19q status. They showed that choline/creatine ratio > 2.4 was associated with the occurrence of anaplastic transformation in patients without 1p/19q codeletion. On the other hand, no anaplastic transformation was observed in patients with 1p/19q codeletion.

In this study, we present a robust and non-invasive method to predict the 1p/19q status of LGG from post-contrast T1- and T2-weighted MR images using convolutional neural networks (CNN).

We use a combination of two commonly acquired image types, T2 and post-contrast T1-weighted images, as input to our classification algorithm. Figure 1 shows examples of LGG image characteristics in post-contrast T1- and T2-weighted images with and without 1p/19q codeletion. Our classification algorithm consists of several pre-processing steps: multi-modal image registration, tumor segmentation, data normalization, and data augmentation. After applying pre-processing steps to data, we use the segmented images to train a multi-scale CNN for prediction of 1p/19q status. A flowchart and implementation details of the multi-scale CNN are shown in Fig. 2.

Figure 3 shows that the multi-scale CNN is overfitting to the original (limited size) training data, when data augmentation is not used, as the distance between the training and validation losses increases over time. The accuracy approaches 100% after 50 epochs for the training set and the loss gets closer to zero, but the performance of trained CNN without data augmentation remains below 80% for the test data (see Table 1). As seen in the Table 1, the performance of CNN with data augmentation (configuration 4) is better than the other configurations without data augmentation.

Data augmentation of original data with k=30-fold (ACC = 89.5) gave the better accuracy on the validation dataset compared to the ones using k=10-fold (ACC = 85.5) and k=20-fold (ACC = 83.3) augmentations. Compared to training on the data without augmentation, accuracy and loss in Fig. 5 fluctuate and are noisier due to variations introduced with data augmentation. The results of the SVM classifier using texture features on the test set were 80% (sensitivity), 82% (specificity), and 81% (accuracy), which are inferior than the results of multi-scale CNN (configuration 4).

Table 2 compares the performance of the best performing configuration using four different optimizers. As seen in Table 2, the influence of optimizers on learning is not much and SGD gives better performance on the test data compared to other optimizers. Figure 4 shows the performance of the best performing CNN configuration 4 on the training data. Fluctuations seen in Fig. 4 are due to randomly introduced augmented data at each epoch. Table 1 shows the performance of multi-scale CNN configurations on the test dataset. Figure 5 shows the validation loss of the best performing CNN (configuration 4) on the validation set. As seen in Figs. 4 and 5, the training and validation losses converge to same value of about 0.2 over time. Table 3 shows the confusion matrix of 1p19q status on test set for the best performing CNN (configuration 4). False negatives are higher than false positives.

In this study, we present a robust and non-invasive method to predict 1p/19q chromosomal arm deletion from post-contrast T1- and T2-weighted MR images using multi-scale CNN approach. As mentioned in previous studies [4, 6‐8], a subset of LGG are sensitive to therapy and have longer progress-free survival while others may progress to HGG. Although histologic grade is the most important factor in therapeutic decision-making, it lacks information about other tumor properties that can impact optimal therapy options and recent reports suggest genomic marker may be more predictive than histologic grade [26]. Adding other information such as 1p/19 chromosomal arms deletion, which has been associated with positive response to therapy, could help improve therapeutic decision-making.

An important challenge in applying deep learning methods to medical images is having an adequate number of data sets. Our multi-scale CNN reached to 100% sensitivity, specificity, and accuracy in training set without data augmentation (see Fig. 3), but had less than 80% specificity and accuracy for the (unseen) test set (see Table 1). This demonstrated that our algorithm did not truly learn the distinctive imaging features, but rather, overfit to the training examples. As seen in Table 1, configuration 4 using augmented data to train our multi-scale CNN approach perform better than the other configurations on the unseen test dataset. Fluctuations as seen in Fig. 4 are results of introducing new augmented training data at each epoch. As seen in the results, augmenting a randomly selected set of the original training data, which is described in the experiment section, at each epoch improves generalization ability of the multi-scale CNN and prevents overfitting. As seen in Table 1, configuration 4 yields the best performance on the test dataset. The performance of configuration 4 in the training and validation datasets is in the same order with its performance in the test dataset. This highlights the generalizability of configuration 4 and its minimal overfitting.

The misclassification rate on the test dataset is about 11%. Some part of this error might be due to the error rate introduced by the FISH test that was used to determine the 1p/19 status. Scheie et al. [13] showed that the reliability of FISH test was 95% and 87.5% for the detection of 1p and 19q deletions, respectively.

As seen in Fig. 5, the loss on validation dataset for configuration 4 stops fluctuating and stabilizes with slight changes after 400 epochs for several epochs. This means that the network is not learning much any more and had converged to a steady state. As seen in Table 2, SGD optimizer gave superior performance on the test dataset compared to other three optimizers.

As seen in our results, deep learning algorithm performs better than the classical machine-learning algorithm using SVM. As mentioned before, Fellah et al. [9] presented the first study that determined the 1p/19q status from DWI, PWI, MR spectroscopy, and conventional MR images. Their misclassification rates were 48 and 40% for using only the conventional MR images and using all multimodal images, respectively. Compared to their results, our best-performing configuration (4) shows better performance in our dataset, with 93.3% sensitivity, 82.2% specificity, and 87.7% accuracy in the unseen test dataset. Moreover, our data was evaluated in a larger dataset (159 LGG vs. 50 LGG). However, it is hard to make the direct comparison since we used different datasets. To the best of our knowledge, our study is the first in using the deep learning to predict 1p19q from MR images in LGGs.

Although the original data size was limited, artificially augmenting data helped increase the volume of our data. It may be that further performance gains will be realized with larger patient populations with more heterogeneous data. Furthermore, including weight regularization such as L1 or L2 in the network might improve generalizability of our networks. Further studies with larger patient populations are required to investigate these and confirm our current findings.

Our multi-scale CNN approach provides promising results for predicting 1p/19q codeletion status non-invasively, based on post-contrast T1 and T2 images. Data augmentation helps counter overfitting and provides learning the invariant patterns in the images. Our presented method could be further improved and potentially used as an alternative to surgical biopsy and pathological analysis for predicting 1p/19q codeletion status.