Automatic Brain Tumor Segmentation Method Based on Modified Convolutional Neural Network.

Area of Science:

• Medical image analysis
• Artificial intelligence in medicine
• Neuro-oncology

Background:

• Brain tumor segmentation is challenging due to severe structural deformation.
• Convolutional neural networks (CNNs) show promise in semantic segmentation tasks.
• Accurate segmentation is crucial for treatment planning and monitoring.

Purpose of the Study:

• To develop a robust CNN algorithm for automatic brain tumor segmentation.
• To improve the accuracy of segmenting enhancing tumor regions.
• To enhance overall brain tumor segmentation performance.

Main Methods:

• A modified U-Net architecture combined with ResNet was developed.
• The BraTS 2017 dataset was utilized for training and testing.
• Weighted cross-entropy loss and data augmentation addressed data imbalance and overfitting.

Main Results:

• Achieved an average Dice score of 0.748 for enhancing tumor segmentation on the validation set.
• Obtained average Dice scores of 0.883 (whole tumor) and 0.781 (tumor core) on the validation set.
• Reported Dice scores of 0.877 (whole tumor), 0.774 (tumor core), and 0.757 (enhancing tumor) on the testing set.

Conclusions:

• The proposed CNN algorithm significantly improves brain tumor segmentation accuracy.
• The modified U-Net with ResNet is effective for segmenting complex brain tumor structures.
• This approach offers a robust solution for automatic brain tumor segmentation in clinical settings.