RAAGR2-Net

RAAGR2-Net performs automated segmentation of brain tumor regions from multimodal MRI using an encoder-decoder deep learning architecture.

Key Features:

• Multimodal MRI Integration: Leverages T1-weighted (T1), T1-weighted with contrast (T1c), T2-weighted (T2), and FLAIR modalities to capture complementary tumor characteristics.
• Encoder-Decoder Architecture: Employs an encoder-decoder backbone for hierarchical feature extraction and pixel-wise segmentation.
• Residual Spatial Pyramid Pooling (RASPP): Uses parallel layers with dilated convolution to capture spatial hierarchies and preserve location information across scales.
• Attention Gate (AG) Module: Emphasizes relevant features from feature maps to focus on tumor regions and suppress background noise.
• Recursive Residual (R2) Block: Recursively refines residual connections to enhance feature extraction and representation depth.
• Data Pre-processing Techniques: Applies N4 bias field correction, z-score normalization, and resampling to the 0–1 range.
• Simultaneous Multi-modality Segmentation: Segments tumor regions across all provided MRI modalities concurrently.
• Reported Performance: Experimental results reported improved fine-grained segmentation relative to existing networks.

Scientific Applications:

• Brain tumor delineation: Segmentation of tumor subregions in MRI for diagnosis, treatment planning, and monitoring therapeutic outcomes in research and clinical contexts.

Methodology:

RAAGR2-Net uses an encoder-decoder network with RASPP (dilated convolutions), Attention Gate modules, and Recursive Residual (R2) blocks; input data are preprocessed with N4 bias field correction, z-score normalization, and resampling to 0–1, and the network is trained and evaluated on the BraTS (Brain Tumor Segmentation) benchmark dataset to assess segmentation across all MRI modalities simultaneously.