BiTr-Unet

BiTr-Unet applies a hybrid convolutional neural network and Transformer architecture to segment whole tumor, tumor core, and enhancing tumor in 3D multi-modal MRI for brain tumor analysis.

Key Features:

• Hybrid Architecture: BiTr-Unet integrates convolutional neural networks (CNNs) and Transformer networks to combine local feature extraction with long-range contextual modeling.
• Local and Global Feature Modeling: CNNs capture spatial hierarchies in 3D medical images while Transformer self-attention models global dependencies.
• Designed for 3D Multi‑modal MRI: The model is tailored for 3D multi-modal MRI scans used in brain tumor segmentation.
• Segmentation Targets: Targets include whole tumor, tumor core, and enhancing tumor segmentation.
• Performance on BraTS2021: On the BraTS2021 validation set it achieved median Dice scores of 0.9335 (whole tumor), 0.9304 (tumor core), and 0.8899 (enhancing tumor) with Hausdorff distances of 2.8284, 2.2361, and 1.4142 respectively; on the testing set Dice scores were 0.9257, 0.9350, and 0.8874 with Hausdorff distances of 3, 2.2361, and 1.4142.

Scientific Applications:

• Brain tumor segmentation: Automated segmentation of brain tumors in 3D multi-modal MRI for research and diagnostic evaluation.
• Tumor component delineation: Precise delineation of whole tumor, tumor core, and enhancing tumor for morphological analysis.
• Clinical research and treatment planning: Provides quantitative segmentation outputs to support clinical research, treatment planning, and monitoring.

Methodology:

The architecture combines convolutional feature extraction and Transformer self-attention to process 3D multi-modal MRI and produce segmentations of whole tumor, tumor core, and enhancing tumor.