3.3 Deep Learning Reconstruction (AI4Recon)

Deep learning reconstruction can be viewed as adding learnable components to analytic/iterative pipelines to improve quality and speed under challenging conditions (noise, sparse sampling, low dose).

---

Three common paradigms

1. Post-processing: analytic reconstruction → neural denoiser/artifact remover
2. Unrolled/learned iterative: unfold optimization (ADMM, primal-dual) into a network
3. End-to-end: learn mapping from measurement domain to image domain

---

Practical caveats

• generalization across scanners/sites
• interpretability and safety (avoid “hallucinated” anatomy)
• clinical validation and standardization

---

Next

Try an end-to-end mini pipeline in Chapter 4 (case studies).