Using Synthetic Training Data for Deep Learning-Based GBM Segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Neuro-oncology Imaging

Background:

• Accurate segmentation of glioblastoma multiforme (GBM) in MRI is critical but challenging due to manual methods being time-consuming and subjective.
• Deep learning offers promising automated solutions for medical image segmentation, addressing limitations of manual delineation.
• Limited availability of annotated medical data, particularly for rare conditions like GBM, hinders the training of deep learning models.

Purpose of the Study:

• To develop and evaluate a fully automatic binary segmentation method for GBMs in 2D MRI.
• To investigate the feasibility of training a segmentation network exclusively on synthetic data.
• To provide an objective, reproducible, and efficient alternative to manual tumor segmentation.

Main Methods:

• A semi-automatic method was employed to generate synthetic GBM data and corresponding ground truth annotations.
• A U-Net-based convolutional neural network architecture was utilized for image segmentation.
• The network was trained exclusively on the synthetically generated dataset.

Main Results:

• The developed U-Net model demonstrated successful segmentation of GBMs.
• Performance was evaluated using real MRI scans, indicating the potential of synthetic data for training.
• The study highlights the effectiveness of deep learning for automated medical image segmentation.

Conclusions:

• Fully automatic segmentation of GBMs is achievable using deep learning trained on synthetic data.
• Synthetic data generation offers a viable solution to overcome data scarcity in medical imaging.
• This approach paves the way for more objective and efficient clinical workflows in neuro-oncology.