DRRNet: Dense Residual Refine Networks for Automatic Brain Tumor Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Glioma is a common and aggressive brain tumor requiring accurate segmentation for effective treatment.
• Manual segmentation of brain tumor MRI is time-consuming and labor-intensive.
• Deep learning methods offer promising solutions for automatic segmentation due to their learning and generalization capabilities.

Purpose of the Study:

• To propose a novel automatic 3D Convolutional Neural Network (CNN) based method for brain tumor segmentation.
• To enhance contextual information capture and improve segmentation performance.
• To reduce computational burden while maintaining high accuracy.

Main Methods:

• A novel 3D CNN architecture based on U-Net, incorporating encoder adaptation blocks and dense connected fusion blocks in the decoder.
• Utilizing a generalized dice loss function to address class imbalance issues.
• Evaluation on the BRATS 2015 testing dataset.

Main Results:

• The proposed model achieved Dice scores of 0.84 for whole tumor, 0.72 for tumor core, and 0.62 for enhancing tumor.
• The method demonstrated accuracy and efficiency comparable to state-of-the-art results.
• Successful segmentation of brain tumors using the developed deep learning approach.

Conclusions:

• The novel 3D CNN-based method provides an accurate and efficient solution for automatic brain tumor segmentation.
• The proposed architecture effectively captures contextual information and handles class imbalance.
• This approach holds significant potential for improving clinical treatment planning and quantitative analysis of gliomas.