ROXSI: Robust Cross-Sequence Semantic Interaction for Brain Tumor Segmentation on Multi-Sequence MR Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning for brain tumor segmentation on multi-sequence MRI shows promise for diagnosis.
• Image noise and artifacts in clinical MRI sequences can significantly degrade segmentation performance.
• Robustness of segmentation models to imaging artifacts is crucial but under-explored.

Purpose of the Study:

• To develop a robust brain tumor segmentation framework mitigating performance loss from multi-sequence MRI noise and artifacts.
• To enhance the reliability of AI-driven diagnostic tools in real-world clinical scenarios.

Main Methods:

• Proposed a novel cross-sequence semantic interaction (CSSI) module leveraging semantic affinity for noise-resilient feature extraction.
• Incorporated batch-level covariance and sequence-level variance regularization mechanisms to suppress background noise and enhance feature representation.
• Evaluated robustness against common artifacts at various perturbation levels and performed blinded clinical evaluation.

Main Results:

• The proposed ROXSI framework demonstrated superior robustness compared to state-of-the-art CNN and Transformer-based models.
• Experimental results on two benchmark datasets confirmed the effectiveness of ROXSI in handling noisy and artifact-affected multi-sequence MRI.
• Clinical evaluation by neuro-radiologists supported the superior performance of ROXSI.

Conclusions:

• ROXSI offers a robust solution for brain tumor segmentation, significantly improving reliability in the presence of common MRI artifacts.
• The framework holds potential for enhancing clinical decision-making in neuro-oncology by providing more dependable segmentation results.
• Further research into robust deep learning models is essential for advancing AI in medical diagnostics.