Brain Tumor Segmentation for Multi-Modal MRI with Missing Information.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Deep convolutional neural networks (DCNNs) show potential for brain tumor segmentation using multi-modal MRI.
• Tumor segmentation is challenged by missing or unusual MRI sequences, necessitating robust models.
• Training separate models for all MRI sequence combinations is impractical.

Purpose of the Study:

• To develop a DCNN framework for robust brain tumor segmentation despite missing MRI sequences.
• To introduce a novel sequence dropout technique to train DCNNs resilient to data variability.
• To evaluate the proposed method's performance on the BraTS 2021 dataset.

Main Methods:

• Implemented a DCNN framework with a sequence dropout technique.
• Trained networks to be robust to missing MRI sequences while utilizing available ones.
• Validated performance on the RSNA-ASNR-MICCAI BraTS 2021 Challenge dataset.

Main Results:

• No significant performance difference was observed with or without sequence dropout when all MRI sequences were available (p > 0.05).
• Sequence dropout significantly improved segmentation performance when key MRI sequences were unavailable.
• For example, Dice Similarity Coefficient (DSC) for enhanced tumor (ET) increased from 0.143 to 0.486 when T1, T2, and FLAIR were used.

Conclusions:

• Sequence dropout is an effective and simple approach for enhancing brain tumor segmentation robustness with missing MRI data.
• The proposed method maintains high performance with complete data while significantly improving it with incomplete data.
• This technique offers a practical solution for clinical settings where MRI data may be variable.