Vox-MMSD: Voxel-wise Multi-scale and Multi-modal Self-Distillation for Self-supervised Brain Tumor Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning for brain tumor segmentation from multi-modal MRI is crucial for diagnosis and treatment.
• Supervised methods struggle with limited labeled data due to costly annotations.
• Self-supervised pre-training offers a solution by learning from unlabeled data.

Purpose of the Study:

• To propose a novel Self-Supervised Learning (SSL) framework for brain tumor segmentation.
• To leverage multi-modal MRI information and multi-scale features effectively.
• To improve segmentation performance, especially with limited labeled data.

Main Methods:

• Developed a Siamese Block-wise Modality Masking (SiaBloMM) strategy for diverse input generation and restoration.
• Introduced Overlapping Random Modality Sampling (ORMS) for multi-scale feature self-distillation.
• Focused on extracting modality-invariant and enhancing voxel-wise representations.

Main Results:

• Achieved a 3.80 percentage point improvement in average Dice score on the BraTS 2024 dataset with limited fine-tuning data.
• Demonstrated improved performance on three other downstream brain tumor datasets, averaging a 3.47 percentage point Dice improvement.
• Outperformed several existing self-supervised learning methods in brain tumor segmentation tasks.

Conclusions:

• The proposed SSL framework effectively utilizes multi-modal and multi-scale features for brain tumor segmentation.
• The method significantly improves segmentation accuracy, particularly in low-data regimes.
• This approach offers a promising direction for developing robust brain tumor segmentation models.