Non-invasive diagnosis of H3 K27M mutant midline glioma

See the article by Su et al. in this issue, pp. 393–401.

The 2016 World Health Organization classification of CNS tumors includes a new subtype: diffuse midline glioma, H3 K27M mutant.¹ These tumors carry a worse prognosis compared with H3 K27M wild-type tumors independent of tumor location.² Since diffuse midline gliomas are commonly located in deep locations, including the brainstem, thalamus, and spinal cord, surgical resection or biopsy can be challenging, with substantial perioperative risks and postoperative morbidities. Thus, a non-invasive method for diagnosing midline glioma with H3 K27M mutations may reduce the need for invasive biopsy and help select patients for targeted therapy.³

On conventional MRI, H3 K27M mutant midline gliomas frequently appear uniform in signal intensity with variable patterns of enhancement.⁴ These tumors can extend to hemispheric locations from midline structures. Hemorrhage and edema are rarely observed. On diffusion-weighted imaging, H3 K27M mutant tumors typically show restricted diffusion⁴ with lower minimal apparent diffusion coefficient value than wild-type tumors.⁵ While these imaging characteristics are helpful for evaluating H3 K27M mutant tumors, there are significant overlaps between mutant and wild-type tumors, limiting their diagnostic utility in selecting the mutant subtype.

Radiomics is a novel method of generating quantitative imaging features based on analysis of pixel intensity configuration or distribution within a lesion of interest or its surrounding tissues. These features commonly include tumor shape, volume, summary statistics of pixel intensity, as well as lesion texture. The features can be individually characterized or combined using machine learning algorithms to construct diagnostic, prognostic, or predictive biomarkers. One important application of the radiomics approach is to non-invasively predict clinically important genetic mutations of tumors. Accurate classification of these molecular markers preoperatively can provide prognostication values and help guide clinical management. For gliomas, a number of radiomic models have been developed to predict isocitrate dehydrogenase 1 and 2,⁶ epidermal growth factor receptor variant III mutations,⁷ and 1p/19q codeletion.⁸ These radiomic models all consist of imaging features that are relatively nondiscriminative when evaluated individually but demonstrate high synergistic classification accuracy when combined using machine learning algorithms.

In the study in this issue by Su and colleagues,⁹ a radiomic model was developed using a retrospective, single institution cohort of 122 patients with newly diagnosed midline gliomas that had confirmed H3 K27M mutational status. The authors extracted radiomic features from tumor regions of interest derived from fluid attenuated inversion recovery (FLAIR) images, and a genetic programming-based automatic machine learning pipeline was applied to these features to generate predictive models. The model search algorithm, the Tree-Based Pipeline Optimization Tool, utilizes a stochastic procedure where multiple machine learning algorithms are evaluated to select the model with the best fit as well as most discriminative features. To avoid overfitting, the authors validated the selected model using a prospectively acquired cohort of 22 patients from the same institution and showed a sensitivity in the validation cohort of 0.8, specificity of 0.917, and accuracy of 0.864. As currently there are no reliable imaging predictors of H3 K27M mutation, the 86% accuracy of the classifying model is a promising result that should invite further investigations to validate this approach.

The most discriminative features that contributed to the accuracy of the final selected model include several shape, first-order, and texture features. Among these features, mean maximum 2D diameter of mutant tumors is shorter than that of wild-type tumors, consistent with a previous analysis of spine H3 K27M mutant gliomas.¹⁰ There are also several other features included in the final model that have been successfully used for prediction of other molecular markers of glioma. Since these radiomic imaging phenotypes may not be directly related to the genotype they intend to predict, further work is needed to investigate how the imaging-based phenotypes discovered in this work are linked to the H3 K27M mutations.

Due to its small sample size and single-institution study design, the current radiomic prediction model needs to be validated on datasets from external sites to demonstrate its generalizability. Moreover, the radiomic model is derived from preselected training data that do not include other potential confounding diagnoses such as nontumor mimickers. Therefore, the current model also needs further evaluation using prospective cases of midline lesions to establish its predictive value. The automatic machine learning pipeline employed in this work was built using open-source software, which should allow the prediction model to be reproduced using different datasets.

The radiomic model developed by Su et al is an important step toward our ability to non-invasively diagnose specific molecular subtypes of CNS tumor. The deep locations of midline gliomas make such prediction model particularly useful clinically. The authors extracted features from FLAIR imaging sequence that is routinely included in standard imaging protocol of brain tumor in most cancer centers. Although a single-sequence input approach simplifies the post-processing workflow and allows the model to be more easily implemented clinically, the performance of the prediction model may further improve with inclusion of other imaging sequences as well as clinical variables.