Automated multi-class high-grade glioma segmentation based on T1Gd and FLAIR images

Abstract

Glioma is the most prevalent primary malignant brain tumor. Segmentation of glioma regions using magnetic resonance imaging (MRI) is essential for treatment planning. However, segmentation of glioma regions is usually based on four MRI modalities, which are T1, T2, T1Gd, and FLAIR. Acquiring these four modalities will increase patients' time inside the scanner and drive up the segmentation process's processing time. Nevertheless, not all these modalities are acquired in some cases due to the limited available time on the MRI scanner or uncooperative patients. Therefore, U-Net-based fully convolutional neural networks were employed for automated segmentation to answer the urgent question: does a smaller number of MRI modalities limit the segmentation accuracy? The proposed approach was trained, validated, and tested on 100 high-grade glioma (HGG) cases twice, once with all MRI sequences and second with only FLAIR and T1Gd. The results on the test set showed that the baseline U-Net model gave a mean Dice score of 0.9166 and 0.9190 on all MRI sequences using FLAIR and T1Gd, respectively. To check for possible performance improvement of the U-Net on FLAIR and T1Gd modalities, an ensemble of pre-trained VGG16, VGG19, and ResNet50 as modified U-Net encoders were employed for automated glioma segmentation based on T1Gd and FLAIR modalities only and compared with the baseline U-Net. The proposed models were trained, validated, and tested on 259 high-grade gliomas (HGG) cases. The results showed that the proposed baseline U-Net model and the ensemble of pre-trained VGG16, VGG19, or ResNet50 as modified U-Net encoders have a mean Dice score of 0.9395, 0.9360, 0.9359, and 0.9356, respectively. The results were also compared to other studies based on four MRI modalities. The work indicates that FLAIR and T1Gd are the most prominent contributors to the segmentation process. The proposed baseline U-Net is robust enough for segmenting HGG sub-tumoral structures and competitive with other state-of-the-art works.

Keywords

• Brain tumor segmentation
• Convolutional neural networks (CNNs)
• HGG
• Pre-trained CNNs
• U-net