Deep learning approaches to non-invasively assess molecular features of gliomas

The recent 2021 World Health Organization classification of central nervous system tumors has reinforced the role of molecular changes in CNS tumor classification, and they increasingly inform the management of brain tumor patients.¹ For example, IDH wild-type status is now required to establish a diagnosis of glioblastoma, and IDH-mutant diffuse astrocytic tumors are a single entity with a numeric (2, 3, or 4) grade. Oligodendrogliomas have been defined by IDH mutation and 1p/19q co-deletion since 2016.²

The identification of key molecular changes, particularly IDH mutation status and 1p/19q co-deletion status is therefore paramount to the classification of CNS tumors and subsequent understanding of prognosis and management of patients. Molecular testing, however, can be a time-consuming process reliant on sufficient surgical tissue. Noninvasive identification of molecular status would be tremendously valuable. Early determination of molecular subgroups could inform clinical trial eligibility or help with decision making of surgical approach or treatment in cases where surgical tissue sampling is not possible.

Prior image analysis approaches have suggested that MRI features may be predictive of IDH mutation and 1p/19q co-deletion status.^3–5 Building on this initial work, deep learning approaches have grown in use in neuro-oncology and leverage neural networks to predict clinically relevant outcomes such as glioma molecular results.⁶ In this study by Clucero et al,⁷ convolutional neural network (CNN) classifiers were trained, validated, and tested with a 3-class or tiered structure to identify IDH mutation and 1p/19q co-deletion status based upon preoperative imaging of glioma patients. The study used a cohort of 384 patients with newly diagnosed gliomas with anatomical imaging and diffusion-weighted imaging (DWI). An external cohort of 147 patients with anatomical imaging from The Cancer Genome Atlas (TCGA) dataset was used to assess the generalization of findings.

Among generated models, 3-class models outperformed tiered models, particularly with respect to overall accuracy in the institutional test set (82% vs 69%). Classifying IDH and 1p/19q co-deletion mutation status simultaneously was advantageous over a stepwise approach of classifying by IDH mutation and then 1p/19q co-deletion. The addition of DWI with apparent diffusion coefficient (ADC) images, which can provide a measure of cellularity, improved model performance. The best model used a 3-class CNN and incorporated ADC images; this demonstrated an overall accuracy of 85.7% (95% CI 0.771-1.0). Of note, even in the best-performing 3-class model that included ADC images, the accuracy of predicting IDH-mutant, 1p/19q codeleted tumors was 60.0% in the testing set, which was lower than accuracies seen in predicting IDH wild-type (95.2%) and IDH-mutant, non-codeleted tumors (88.9%).

Among limitations, there was a small sample size of each individual molecular subgroup, and further validation of models in external datasets is warranted. While the authors should be commended for a study design that incorporated an external dataset (TCGA), it was composed of a different distribution of molecular subgroups than the original dataset. While the value of possibly integrating DWI is intriguing, ADC images were not available in the TCGA imaging and therefore could not be validated in this external dataset.

An important critique of deep learning approaches has been the difficulty in understanding the workings of the “black box” that leads to model-generated outputs and predictions. Machine learning approaches often can rank important features, but there is not an obvious means to do this with deep learning approaches. The authors address this by using GradCAM, a heatmap-based feature attribution method. By employing this technique, the authors were able to confirm that the models were extracting features from relevant areas of images (ie, tumor regions). The use of GradCAM analysis provides reassuring visualization of the generation of models, though further work is needed to understand how imaging features are related to the molecular status of glioma.

While previous work has used deep learning techniques to predict IDH mutation status,^8,9 there has been comparatively less published work available on the prediction of 1p/19 co-deletion status.¹⁰ Given the distinct biology and treatment sensitivity conferred to patients with IDH mutation and 1p/19q co-deletion, differentiating these tumors from IDH-mutant non-codeleted tumors and IDH wild-type tumors is clinically meaningful. Oligodendrogliomas frequently contain calcifications, and future research should consider the possible benefit of using CT imaging and susceptibility-weighted MRI, in addition to anatomical imaging and DWI, to better predict 1p/19 co-deletion. Although the accuracy of predicting IDH-mutant 1p/19q codeleted patients was suboptimal in this study, the work by Clucero et al represents an important step in developing methods to non-invasively identify clinically relevant molecular subgroups of gliomas.

Funding

There were no funding sources for this work.

Conflict of interest statement. None.