Modality-Pairing Learning for Brain Tumor Segmentation

Automatic brain tumor segmentation from multi-modality Magnetic Resonance Images (MRI) using deep learning methods plays an important role in assisting the diagnosis and treatment of tumor. However, previous mostly ignore latent relationship among different modalities. In this work, we propose a novel end-to-end Modality-Pairing method for segmentation. Paralleled branches are designed to exploit modality features series layer connections utilized capture complex relationships abundant information We also use consistency loss minimize prediction variance between two branches. Besides, rate warmup strategy is adopted solve problem training instability early over-fitting. Lastly, average ensemble multiple models some post-processing techniques get final results. Our tested on BraTS 2020 online testing dataset, obtaining promising performance, with dice scores 0.891, 0.842, 0.816 whole tumor, core enhancing respectively. won second place Challenge task.