Lifelong learning enabled GAN framework for brain MRI FLAIR synthesis and tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate synthesis of missing MRI modalities is vital for multi-modal neuro-imaging.
• FLAIR MRI sequences are crucial for visualizing tumor boundaries and edema, but are often unavailable or of low quality.
• This limitation impacts diagnostic pipelines in real-world clinical settings.

Purpose of the Study:

• To develop a 2.5D Generative Adversarial Network (GAN)-based framework for synthesizing missing FLAIR MRI sequences.
• To integrate this synthesis framework with a tumor segmentation model within a lifelong learning strategy.
• To evaluate the model's performance on sequential tasks using the BraTS2020 dataset.

Main Methods:

• A 2.5D GAN framework using a U-Net generator with residual blocks and CBAM, and a PatchGAN discriminator.
• A U-Net-based segmentor for downstream utility.
• Training incorporated perceptual, structural losses, and segmentation consistency constraints within a lifelong learning strategy (EWC) across 4 sequential tasks.

Main Results:

• Achieved high quantitative metrics for synthesis: average SSIM of 0.902, PSNR of 28.375 dB, MAE of 0.023, and LPIPS of 0.119.
• Demonstrated strong segmentation performance with an average Dice score of 0.79 and HD95 of 7.68 mm.
• Showcased effective backward and forward knowledge transfer, confirming pipeline efficacy and lifelong learning potential.

Conclusions:

• The proposed synthesis-segmentation pipeline effectively generates missing FLAIR MRI sequences and performs accurate tumor segmentation.
• The lifelong learning approach enables adaptation to new data distributions with minimal catastrophic forgetting.
• This framework holds significant potential for improving medical image synthesis and segmentation in clinical practice.