Improving brain tumor segmentation with anatomical prior-informed pre-training.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Accurate glioblastoma segmentation in multi-parameter MRI is crucial for neurosurgery and treatment.
• Transformer models show potential but require extensive annotated data.
• Current self-supervised learning methods for brain tumor segmentation lack anatomical prior integration.

Purpose of the Study:

• To develop an anatomical prior-informed masking strategy for enhancing masked autoencoder pre-training.
• To improve the robustness and data efficiency of brain tumor segmentation models.
• To integrate anatomical knowledge with data-driven reconstruction for better segmentation.

Main Methods:

• Proposed an anatomical prior-informed masking strategy for masked autoencoders.
• Investigated tumor presence likelihood in brain structures to guide masking.
• Combined data-driven reconstruction with anatomical knowledge for pre-training.

Main Results:

• The proposed method significantly improved brain tumor segmentation performance on the BraTS21 dataset.
• Outperformed state-of-the-art self-supervised learning techniques.
• Demonstrated enhanced accuracy and data efficiency compared to random masking.

Conclusions:

• Integrating anatomical priors into self-supervised learning enhances computational efficiency and segmentation precision.
• The proposed method offers a promising approach for improving glioblastoma delineation.
• Further integration of anatomical priors and vision approaches holds significant potential.