Cross-Modal Distillation to Improve MRI-Based Brain Tumor Segmentation With Missing MRI Sequences

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Convolutional neural networks (CNNs) for brain tumor segmentation typically require complete MRI sequences for training and inference.
• Clinical settings often present incomplete MRI data, limiting the applicability of standard CNN models.

Purpose of the Study:

• To develop a cross-modal distillation approach for brain tumor segmentation.
• To enhance CNN models for inference using only single-sequence MRI data, outperforming single-sequence models.

Main Methods:

• Proposed a cross-modal distillation technique leveraging multi-sequence MRI data during training.
• Evaluated the enriched CNN model on BraTS 2018 and in-house datasets for whole tumor and tumor core segmentation.
• Inference was performed using only T1-weighted MRI sequences.

Main Results:

• Cross-modal distillation significantly improved Dice scores for both whole tumor and tumor core segmentation with single-sequence inference.
• On an in-house dataset, the method achieved 79.04% (whole tumor) and 69.39% (tumor core) Dice scores.
• Outperformed a standard single-sequence U-Net model (73.60% and 62.62% Dice scores, respectively).

Conclusions:

• Cross-modal distillation is an effective strategy to improve single-sequence CNN models for brain tumor segmentation.
• The approach mitigates performance compromises caused by missing MRI sequences.
• Enhances the clinical utility of CNNs in scenarios with limited MRI data availability.