Conditional Invertible Neural Networks for Medical Imaging

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Science

Background:

• Deep learning is increasingly used for inverse problems, but often provides only point estimates.
• Quantifying uncertainty is crucial for ill-posed inverse problems in medical imaging.
• Existing deep learning methods for medical imaging often lack uncertainty estimation.

Purpose of the Study:

• To apply generative flow-based models with invertible neural networks to medical imaging inverse problems.
• To evaluate different invertible neural network architectures.
• To investigate the impact of base distribution choice on reconstruction quality.

Main Methods:

• Generative flow-based models utilizing invertible neural networks.
• Application to low-dose computed tomography (CT) and accelerated medical resonance imaging (MRI).
• Testing various invertible neural network architectures and conducting ablation studies.

Main Results:

• Invertible neural networks show promise for medical imaging reconstruction.
• The choice of base distribution significantly impacts reconstruction quality.
• Radial distributions outperform standard Gaussian distributions for these tasks.

Conclusions:

• Generative flow-based models are effective for uncertainty-aware medical image reconstruction.
• Optimizing the base distribution is key to improving reconstruction performance.
• This approach offers a pathway to more robust and reliable medical imaging analysis.