ResSAXU-Net for multimodal brain tumor segmentation from brain MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Glioma is the most common and deadliest brain tumor, necessitating accurate classification via magnetic resonance imaging (MRI) for effective treatment.
• Current automated brain tumor classification methods, particularly U-Net models, face challenges with feature extraction and information loss during processing.
• Standardized and accurate classification is crucial due to glioma's aggressive nature and varied characteristics.

Purpose of the Study:

• To enhance the accuracy of U-Net models for brain tumor classification.
• To address limitations in feature extraction and information loss in existing U-Net architectures.
• To develop an improved model, Ressaxu-Net, for precise glioma classification from MRI data.

Main Methods:

• Developed Ressaxu-Net, an enhanced U-Net model incorporating deep residual networks and squeeze-excitation networks.
• Utilized deep residual networks to improve feature information extraction and mitigate network damage.
• Integrated squeeze-excitation networks to prioritize essential feature maps and reduce information loss.
• Implemented a fusion loss function combining dice loss and cross-entropy to handle network convergence and data imbalance.

Main Results:

• Ressaxu-Net demonstrated high performance on the Brats2018 and Brats2019 datasets.
• Achieved Dice Similarity Coefficients of 0.9597 (total tumor), 0.9618 (intratumoral), and 0.9595 (elevated tumor).
• Showed significant improvements of 8.10%, 15.88%, and 17.33% in classifying different tumor components.

Conclusions:

• Ressaxu-Net effectively enhances brain tumor classification accuracy compared to standard U-Net models.
• The model's architecture successfully addresses feature extraction and information loss issues.
• Ressaxu-Net shows competitive effectiveness in accurately classifying multiple brain tumors from MRI data.