A Two-Stage Cascade Model with Variational Autoencoders and Attention Gates for MRI Brain Tumor Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate brain tumor segmentation from MRI is crucial for clinical decision-making.
• Existing methods face challenges in precise subregional segmentation and overfitting.

Purpose of the Study:

• To develop an advanced two-stage model for automatic brain tumor subregional segmentation.
• To improve the accuracy and robustness of segmentation using deep learning techniques.

Main Methods:

• A two-stage encoder-decoder network incorporating variational autoencoder (VAE) regularization.
• Attention gates are integrated into the second stage, trained with augmented data from the first stage.
• The model was evaluated on the BraTS 2020 validation and testing datasets.

Main Results:

• Achieved high Dice scores (e.g., 0.9041 for whole tumor on validation set) and low Hausdorff distances (e.g., 4.953 on validation set).
• Demonstrated strong performance on the BraTS 2020 testing dataset with Dice scores of 0.8729, 0.8357, and 0.8205.
• VAE regularization and attention gates effectively addressed overfitting and improved segmentation accuracy.

Conclusions:

• The proposed two-stage model with VAE regularization and attention gates offers a robust solution for brain tumor subregional segmentation.
• The method shows significant potential for improving brain tumor diagnosis, monitoring, and treatment planning.
• Public availability of the code facilitates further research and development in the field.