Chapter 23: Self-Supervised and Unsupervised Methods

Learning Objectives

• Implement Deep Image Prior (DIP) and Deep Decoder for untrained reconstruction from a single measurement
• Apply Noise2Noise, Noise2Self, and Noise2Void for training without clean ground truth
• Derive Stein's Unbiased Risk Estimate (SURE) as a surrogate MSE loss and use it for unsupervised denoiser training
• Implement equivariant imaging using known symmetry groups and measurement consistency
• Evaluate foundation models and transfer learning strategies for RF imaging with limited data