3D convolutional neural networks for tumor segmentation using long-range 2D context.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Convolutional Neural Networks (CNNs) excel in medical image analysis.
• Processing large MRI volumes with CNNs often involves subvolume analysis (slices or 3D patches).
• Existing methods face challenges with computational cost and missing MR sequences.

Purpose of the Study:

• To develop an efficient deep learning model for multisequence MR image tumor segmentation.
• To combine short-range 3D and long-range 2D contextual information effectively.
• To enhance robustness against missing MR sequences during training.

Main Methods:

• Introduced a CNN-based model integrating 3D and 2D contexts.
• Proposed modality-specific subnetworks for improved robustness.
• Implemented a hierarchical decision process to combine multiple segmentation models.
• Developed an efficient algorithm for training large CNN models.

Main Results:

• Evaluated on the BRATS 2017 challenge for multiclass malignant brain tumor segmentation.
• Achieved high performance with median Dice scores: 0.918 (whole tumor), 0.883 (tumor core), and 0.854 (enhancing core).
• Generated accurate tumor segmentations.

Conclusions:

• The proposed deep learning approach offers an efficient and robust solution for brain tumor segmentation.
• The model effectively leverages multimodal MRI data.
• The method demonstrates state-of-the-art performance on a challenging benchmark dataset.