Nested Dilation Networks for Brain Tumor Segmentation Based on Magnetic Resonance Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Brain tumors are a leading cause of cancer mortality globally.
• Manual tumor segmentation is time-consuming and relies heavily on clinician expertise.
• Accurate, automated segmentation is crucial for effective brain tumor diagnosis and treatment planning.

Purpose of the Study:

• To develop and evaluate an advanced deep learning algorithm for accurate 3D multimodal brain tumor segmentation.
• To improve upon existing U-Net architectures for enhanced feature extraction and segmentation performance.
• To investigate the efficacy of a novel cascade training strategy for multi-class brain tumor segmentation.

Main Methods:

• Proposed Nested Dilation Networks (NDNs), a U-Net-inspired convolutional neural network (CNN).
• Incorporated residual blocks nested with dilations (RnD) and squeeze-and-excitation (SE) blocks for feature enrichment.
• Employed a cascade training strategy, decomposing the task into sequential binary segmentation problems.
• Utilized various data augmentation techniques and explored different loss functions to address class imbalance.

Main Results:

• NDNs achieved improved Dice similarity scores compared to standard U-Net and its variants.
• Single-pass training with Dice loss yielded scores of 0.6652 (edema), 0.5880 (non-enhancing), and 0.6682 (enhancing) tumors.
• Cascade training further boosted scores to 0.7043 (edema), 0.5889 (non-enhancing), and 0.7206 (enhancing) tumors.

Conclusions:

• The proposed NDNs with cascade training demonstrate superior performance for brain tumor segmentation over existing methods.
• This deep learning approach offers a reliable and efficient solution for automatic and accurate tumor segmentation.
• The study provides valuable insights for advancing automated diagnostic tools in neuro-oncology.