A deep-learning framework for brain tumor segmentation via three-dimensional mass-preserving geometric transformation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate brain tumor segmentation is crucial for diagnosis and treatment planning.
• Current deep learning models often struggle with the irregular shapes and variability of brain tumors in MR images.

Purpose of the Study:

• To develop a robust and efficient deep learning framework for brain tumor segmentation.
• To improve the accuracy and efficiency of brain tumor segmentation using novel image transformation and segmentation strategies.

Main Methods:

• A novel three-dimensional (3D) mass-preserving geometric transformation (MPGT) using a homotopy method to standardize irregular brain MR images.
• A modified two-phase segmentation strategy and a postprocessing technique to minimize inference time and enhance lesion-wise performance.
• Integration of the proposed framework with the nnU-Net model for validation.

Main Results:

• The framework achieved high Dice scores on the BraTS Challenge 2023 dataset: 0.9282 (Whole Tumor), 0.8812 (Tumor Core), and 0.8527 (Enhanced Tumor).
• Performance was competitive with and superior to top-ranking entries in the BraTS 2023 competition.
• The MPGT effectively preserved local mass ratios and global structural integrity for structured deep learning input.

Conclusions:

• The proposed deep learning framework offers a robust and efficient solution for brain tumor segmentation.
• The novel MPGT and optimized segmentation strategy significantly enhance segmentation accuracy and performance.
• This approach shows great promise for clinical applications in neuro-oncology.