Improving the noninvasive classification of glioma genetic subtype with deep learning and diffusion-weighted imaging

Abstract

Background: Diagnostic classification of diffuse gliomas now requires an assessment of molecular features, often including IDH-mutation and 1p19q-codeletion status. Because genetic testing requires an invasive process, an alternative noninvasive approach is attractive, particularly if resection is not recommended. The goal of this study was to evaluate the effects of training strategy and incorporation of biologically relevant images on predicting genetic subtypes with deep learning.

Methods: Our dataset consisted of 384 patients with newly diagnosed gliomas who underwent preoperative MRI with standard anatomical and diffusion-weighted imaging, and 147 patients from an external cohort with anatomical imaging. Using tissue samples acquired during surgery, each glioma was classified into IDH-wildtype (IDHwt), IDH-mutant/1p19q-noncodeleted (IDHmut-intact), and IDH-mutant/1p19q-codeleted (IDHmut-codel) subgroups. After optimizing training parameters, top performing convolutional neural network (CNN) classifiers were trained, validated, and tested using combinations of anatomical and diffusion MRI with either a 3-class or tiered structure. Generalization to an external cohort was assessed using anatomical imaging models.

Results: The best model used a 3-class CNN containing diffusion-weighted imaging as an input, achieving 85.7% (95% CI: [77.1, 100]) overall test accuracy and correctly classifying 95.2%, 88.9%, 60.0% of the IDHwt, IDHmut-intact, and IDHmut-codel tumors. In general, 3-class models outperformed tiered approaches by 13.5%-17.5%, and models that included diffusion-weighted imaging were 5%-8.8% more accurate than those that used only anatomical imaging.

Conclusion: Training a classifier to predict both IDH-mutation and 1p19q-codeletion status outperformed a tiered structure that first predicted IDH-mutation, then 1p19q-codeletion. Including apparent diffusion coefficient (ADC), a surrogate marker of cellularity, more accurately captured differences between subgroups.

Keywords: ADC; convolutional neural network; deep learning; diffusion-weighted imaging; glioma subtype.

Fig. 1

Schematic of image processing strategy. (A) Segmented contrast-enhancing (CEL) or T2 (T2L) lesions were used to automatically select the slices by first selecting the central slice with maximum area and expanding every 5 mm until the boundary of the lesion was reached. (B) This process was repeated in each direction: axial, coronal, and sagittal. (C) Images were then cropped to either the brain mask, the T2L, a standard size, or were not cropped. (D) Three of the four sequences of interest (T2-FLAIR, T1c, T2, and ADC) were placed in the R, G, and B channels of a color image that was used as the input to the network.

Fig. 2

Main insights from the hyperparameter search and effects of including ADC as an input image. (A–C) Hyperparameter search: (A) A slice-by-slice prediction approach improved the ability to achieve high accuracy on the validation data compared to average pooling of slices. (B) Cropping to the T2-lesion reduced the validation accuracy of the 3-class model. (C) Pre-training increased validation accuracy of the model. (D–F) Benefit of including ADC as an input image: (D) For the 3-class models, lower generalization accuracy was observed when ADC replaced T1c, while replacing the T2 image with ADC achieved the best performance. (E) For the IDH-mutation tier, test accuracy was slightly improved using T1c, T2-FLAIR, and ADC as input images. (F) For the 1p19q mutation tier, replacing T1c with ADC significantly improved test accuracy.

Fig. 3

Final patient accuracy and class accuracies of the final four models: (A) A 2-tiered structure with all anatomical images (left) and ADC replacing the T1c image in the second tier (right). (B) A 3-class structure with anatomical images only (left) and ADC replacing T2 as the third channel (right). An increase in the IDHmut-codel accuracy and overall accuracy was observed when ADC maps were included as input images. The best model was the 3-class model that included ADC.

Fig. 4

Visualization of imaging features and GradCAM analysis of the best (A) and worst (B) predicted patients with the final 3-class model that included ADC. GradCAM maps of the worst predicted patients often indicated that these models were looking outside of the tumor region when making their decision (white ellipses), compared to looking at the lesion for all of the correctly predicted patients. Top column lists true groupings. IDHwt and IDHmut-intact were misclassified as IDHmut-codel gliomas, while the IDHmut-codel examples were predicted as IDHmut-intact.