Unsupervised brain MRI tumour segmentation via two-stage image synthesis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Deep learning excels at automated brain tumor segmentation but requires extensive expert annotations, which are costly and time-consuming.
• Unsupervised learning using synthetic data presents a promising alternative, yet the domain gap between real and synthetic data hinders accuracy.

Purpose of the Study:

• To develop an unsupervised approach for brain tumor segmentation on magnetic resonance (MR) images by addressing the domain gap using a two-stage image synthesis strategy.
• To generate realistic synthetic brain tumor data for training deep learning models, thereby improving segmentation performance without relying on manual annotations.

Main Methods:

• A two-stage image synthesis strategy is proposed: first, a junior segmentation model is trained on synthetic data and used to generate pseudo-labels for real images.
• Second, realistic synthetic images are created by combining real brain images with these pseudo-labels to train a senior segmentation model.

Main Results:

• The proposed SynthTumour approach effectively bridges the domain gap between real and synthetic MR images.
• SynthTumour outperforms existing unsupervised methods in brain tumor segmentation and demonstrates high performance in ischemic stroke lesion segmentation across five datasets.

Conclusions:

• The developed two-stage synthetic data generation strategy enables accurate unsupervised brain tumor segmentation.
• SynthTumour offers a viable and high-performing alternative to supervised methods, reducing the need for expert annotations in medical image analysis.