[Paper Review] Fully Automatic Binary Glioma Grading based on Pre-Therapy MRI using 3D Convolutional Neural Networks

Milan Decuyper, Roel Van Holen|arXiv (Cornell University)|Aug 5, 2019

Brain Tumor Detection and Classification7 references4 citations

TL;DR

This study presents a fully automatic 3D CNN-based system for binary glioma grading using pre-therapy MRI, combining a 3D U-Net for tumor segmentation and a 3D residual network for classification. It achieves 91% accuracy on a heterogeneous dataset of 549 patients and 92% on independent hospital data, demonstrating robust performance across diverse imaging protocols.

ABSTRACT

The optimal treatment strategy of newly diagnosed glioma is strongly influenced by tumour malignancy. Manual non-invasive grading based on MRI is not always accurate and biopsies to verify diagnosis negatively impact overall survival. In this paper, we propose a fully automatic 3D computer-aided diagnosis (CAD) system to non-invasively differentiate high-grade glioblastoma from lower-grade glioma. The approach consists of an automatic segmentation step to extract the tumour ROI followed by classification using a 3D convolutional neural network. Segmentation was performed using a 3D U-Net achieving a dice score of 88.53% which matches top performing algorithms in the BraTS 2018 challenge. The classification network was trained and evaluated on a large heterogeneous dataset of 549 patients reaching an accuracy of 91%. Additionally, the CAD system was evaluated on data from the Ghent University Hospital and achieved an accuracy of 92% which shows that the algorithm is robust to data from different centres.

Motivation & Objective

• To develop a fully automatic, non-invasive method for differentiating high-grade glioblastoma (GBM) from lower-grade glioma (LGG) using pre-therapy MRI.
• To overcome the limitations of manual MRI-based grading, which suffers from low sensitivity (55–83%) and high false-negative rates due to non-enhancing high-grade tumors.
• To reduce reliance on invasive biopsies, which negatively impact patient survival, by enabling accurate non-invasive diagnosis.
• To ensure robustness across diverse imaging protocols and clinical centers by training on a large, heterogeneous dataset from multiple sources.
• To achieve state-of-the-art performance in glioma grading using a two-stage pipeline combining 3D U-Net segmentation and 3D CNN classification.

Proposed method

• A 3D U-Net with instance normalization and leaky ReLU activations was trained on 285 BraTS 2018 patients to segment whole tumor volumes using T1, T1ce, T2, and FLAIR sequences.
• During training, random masking of T1 or T2/FLAIR channels improved robustness to missing MRI modalities in clinical practice.
• The segmentation model used soft dice loss with L2 weight decay and ADAM optimization on 128×128×128 patches with a batch size of 2.
• Tumor regions of interest (ROIs) were extracted from T1ce scans, resized to 112×112×112, and fed into a 3D residual classification network with one 7×7 kernel convolution followed by four residual blocks.
• The classification network used SGD with momentum, L2 weight decay, and on-the-fly data augmentation (flipping, rotation, padding) to prevent overfitting.
• The system was trained on 400 patients (200 GBM, 200 LGG), validated on 69, and tested on 80 TCIA patients and 112 from Ghent University Hospital, with no overlap between training and test data for segmentation and classification.

Experimental results

Research questions

• RQ1 Can a fully automatic 3D CNN pipeline achieve high accuracy in differentiating high-grade glioblastoma from lower-grade glioma using only pre-therapy MRI?
• RQ2 How does the system perform on independent, multi-center data from different imaging protocols?
• RQ3 To what extent does the use of data augmentation and modality masking improve robustness to missing MRI sequences?
• RQ4 Does the two-stage approach—segmentation followed by 3D CNN classification—achieve state-of-the-art performance compared to existing CAD systems?
• RQ5 Can the model generalize effectively to real-world clinical data beyond benchmark datasets like BraTS?

Key findings

• The 3D U-Net achieved a dice score of 88.53% on the BraTS 2018 test set when using all four MRI sequences, matching state-of-the-art performance.
• The system reached 91.25% accuracy, 96.29% AUC, and 82.19% MCC on the TCIA test set, demonstrating high diagnostic performance.
• On independent data from Ghent University Hospital, the system achieved 91.96% accuracy, 93.39% AUC, and 83.91% MCC, confirming robustness to cross-center variations.
• Sensitivity was 91.18% on TCIA and 90.48% on hospital data, indicating strong detection of high-grade glioblastoma cases.
• Specificity reached 91.30% on TCIA and 93.88% on hospital data, showing high precision in identifying lower-grade gliomas.
• The model maintained high performance even when only T1ce and FLAIR or T1ce and T2 sequences were available, indicating strong generalization to incomplete data.