Deep Learning Methodology for Differentiating Glioma Recurrence From Radiation Necrosis Using Multimodal Magnetic Resonance Imaging: Algorithm Development and Validation

Abstract

Background: The radiological differential diagnosis between tumor recurrence and radiation-induced necrosis (ie, pseudoprogression) is of paramount importance in the management of glioma patients.

Objective: This research aims to develop a deep learning methodology for automated differentiation of tumor recurrence from radiation necrosis based on routine magnetic resonance imaging (MRI) scans.

Methods: In this retrospective study, 146 patients who underwent radiation therapy after glioma resection and presented with suspected recurrent lesions at the follow-up MRI examination were selected for analysis. Routine MRI scans were acquired from each patient, including T1, T2, and gadolinium-contrast-enhanced T1 sequences. Of those cases, 96 (65.8%) were confirmed as glioma recurrence on postsurgical pathological examination, while 50 (34.2%) were diagnosed as necrosis. A light-weighted deep neural network (DNN) (ie, efficient radionecrosis neural network [ERN-Net]) was proposed to learn radiological features of gliomas and necrosis from MRI scans. Sensitivity, specificity, accuracy, and area under the curve (AUC) were used to evaluate performance of the model in both image-wise and subject-wise classifications. Preoperative diagnostic performance of the model was also compared to that of the state-of-the-art DNN models and five experienced neurosurgeons.

Results: DNN models based on multimodal MRI outperformed single-modal models. ERN-Net achieved the highest AUC in both image-wise (0.915) and subject-wise (0.958) classification tasks. The evaluated DNN models achieved an average sensitivity of 0.947 (SD 0.033), specificity of 0.817 (SD 0.075), and accuracy of 0.903 (SD 0.026), which were significantly better than the tested neurosurgeons (P=.02 in sensitivity and P<.001 in specificity and accuracy).

Conclusions: Deep learning offers a useful computational tool for the differential diagnosis between recurrent gliomas and necrosis. The proposed ERN-Net model, a simple and effective DNN model, achieved excellent performance on routine MRI scans and showed a high clinical applicability.

Keywords: deep learning; multimodal MRI; progression; pseudoprogression; radiation necrosis; recurrent tumor.

Figure 1

The T1, T2, and T1c magnetic resonance imaging (MRI) sequences of 4 patients with their histograms of the voxels within the lesion masks. Patients (a) and (b) represent recurrent tumors; patients (c) and (d) represent radionecrosis lesions. The lesion masks were manually drawn using the software ITK-SNAP, generally used for delineating regions of interest. The histograms were created for individual sequences and further smoothed using the Hann filter. ITK: Insight Toolkit. T1: T1-weighted MRI; T1c: gadolinium-contrast-enhanced T1-weighted MRI; T2: T2-weighted MRI.

Figure 2

The selection process for the patient cohorts in this study. MRI: magnetic resonance imaging; T1: T1-weighted MRI; T1c: gadolinium-contrast-enhanced T1-weighted MRI; T2: T2-weighted MRI.

Figure 3

Overview of the proposed approach. (a) The co-registered multimodal images were fused as a multichannel RGB image with T1, T2, and T1c images representing the Red, Green and Blue channels, respectively. (b) The multichannel magnetic resonance (MR) images were used to train the deep neural network (DNN) models that classified the test MR images as either a recurrent tumor or radiation necrosis. (c) Architecture of the proposed efficient radionecrosis neural network (ERN-Net). ReLU: rectified linear unit; T1: T1-weighted magnetic resonance imaging (MRI); T1c: gadolinium-contrast-enhanced T1-weighted MRI; T2: T2-weighted MRI.

Figure 4

Plots showing (a) performance of the deep neural network (DNN) models on multimodal magnetic resonance imaging in the image-based classification task and (b) performance of the DNN models and neurosurgeons in the subject-based classification task. Performance of the DNN models was evaluated using the area under the curve (AUC) of the receiver operating characteristic curves, while the five neurosurgeons’ sensitivity and specificity scores are represented by the red dots. ERN-Net: efficient radionecrosis neural network; ResNet: residual neural network; VGG: Visual Geometry Group.

Figure 5

A T1c tumor image and its corresponding tumor masks created by two neuroradiologists independently, which shows the disagreement between annotators. MRI: magnetic resonance imaging; T1c: gadolinium-contrast-enhanced T1-weighted MRI.