D2-Net

D2-Net performs brain tumor segmentation from incomplete multi-modal Magnetic Resonance Imaging (MRI) by disentangling modality-specific and tumor-specific representations.

Key Features:

• Modality Disentanglement Stage (MD-Stage): Incorporates a spatial-frequency joint modality contrastive learning scheme to decouple modality-specific information from multi-modal MRI.
• Tumor-Region Disentanglement Stage (TD-Stage): Employs an affinity-guided dense tumor-region knowledge distillation mechanism that aligns features from a disentangled binary teacher network with a holistic student network.
• Missing-modality robustness: Explicitly captures correlations among modality-specific information and tumor-specific knowledge to enable segmentation when some MRI modalities are missing.
• Benchmark evaluation: Demonstrated superior segmentation performance over state-of-the-art methods on the BraTS-2018 dataset under missing-modality conditions.

Scientific Applications:

• Clinical brain tumor segmentation: Automatic segmentation of brain tumors from multi-modal MRI to support diagnosis and prognosis when one or more modalities are unavailable.
• Algorithm development and benchmarking: Development and evaluation of segmentation methods for incomplete multi-modal MRI using public datasets such as BraTS-2018.

Methodology:

Spatial-frequency joint modality contrastive learning in a Modality Disentanglement Stage and an affinity-guided dense tumor-region knowledge distillation in a Tumor-Region Disentanglement Stage that aligns disentangled binary teacher features with a holistic student to decompose tumor-specific representations and extract discriminative holistic features.