MRI pulse sequence integration for deep-learning-based brain metastases segmentation.

Area of Science:

• Medical imaging analysis
• Deep learning in radiology
• Neuro-oncology imaging

Background:

• Magnetic resonance (MR) imaging is crucial for diagnosing conditions like brain metastases.
• Deep learning shows promise for automated medical image segmentation.
• Integrating multiple MR pulse sequences can enhance diagnostic accuracy, but optimal methods are unclear.

Purpose of the Study:

• To evaluate architectural features for integrating MR pulse sequences in metastasis segmentation.
• To develop a robust deep learning method for segmentation using any subset of available pulse sequences.
• To address the limitation of networks failing when input pulse sequences differ from training data.

Main Methods:

• Utilized a 2.5D DeepLabv3 segmentation network with four MR pulse sequence inputs.
• Investigated early, middle, and late feature integration schemas.
• Explored various weight-sharing modes for parallel network branches.
• Introduced a novel integration level dropout layer for robustness to missing input sequences.

Main Results:

• Integration strategies and weight sharing favoring low variance performed best with limited training data (n=100).
• Input-level dropout preserved network performance while enabling inference with missing pulse sequences.
• Demonstrated network generalizability and robustness on data from an external center.
• Network visualization identified key input features for performance.

Conclusions:

• Developed a framework for robust deep learning networks in medical imaging.
• Achieved enhanced robustness to missing data while maintaining performance.
• Provides a method for more reliable computer-aided diagnosis in MR imaging.