Click here to

Session: Imaging: MRI Image Analysis, Biomarkers, and Radiomics [Return to Session]

3D Brain Tumor Segmentation Through Multimodal Weighted Network in MRI

Z Zhou1*, R Wang2, J Yang3, J Guo4, (1) University of Central Missouri, Warrensburg, MISSOURI, (2) Xidian University, ,,CN, (3) Xidian University, Xi'an, ,CN, (4) Xidian University,


WE-IePD-TRACK 2-3 (Wednesday, 7/28/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: Accurate brain tumor segmentation in Magnetic resonance imaging (MRI) images plays an important role in image-guided surgery and treatment planning, while unreliable may cause mislead surgery. We aim to develop a new multimodal weighted network (MW-Net) for three-dimension (3D) brain tumor segmentation.

Methods: Totally 285 patients including 210 high grade and 75 low grade gliomas from Brats 2018 dataset are used in this study. The registration is performed on four available modalities such as T1, T1Gd, T2 and FLAIR on each case. In MW-Net, 3D-U-Net which is developed based on U-net are used as base model, which consists of convolutional, pooling and upsampling layers in 3D way. The activation function is ReLU, and the loss function is cross entropy. Adam function is used to tune the model parameters. Then four modalities with the corresponding weights fed into the shared 3D-U-Net model. A backpropagation algorithm is employed to train MW-Net. In testing stage, the 3D brain image is directly fed into trained MW-Net and the 3D segmentation results can be obtained. Dice coefficient (Dice), Positive Predictive Value (PPV) are sensitivity (SEN) are used as evaluation criterion. The better segmentation results can be obtained with higher Dice, SEN and lower PPV.

Results: Five-fold cross validation is performed in this study. The evaluation results for MW-Net on Dice, PPV and SEN are 0.8824, 0.8823, and 0.8963, respectively, while 3D-U-Net are 0.8524, 0.8995, and 0.8475.

Conclusion: A new multimodal weighted network (MW-Net) for three-dimension (3D) brain tumor segmentation was developed. The experimental results demonstrated that MW-Net can obtain better segmentation performance compared with available methods.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email