Overview
Syllabus
Intro
Magnetic Resonance Imaging
Data Pipeline
Raw Data in MRI
k-space Sampling
Constraints
Can We Speed Things Up??
MRI Super-Resolution Convert low resolution (LR) to high resolution (HR)
Domain Knowledge • Embedding physical principles into the model
Reducing Data Requirements Domain knowledge reduce extent of ill-posed inverse problems
What if we have no training data
Deep Image Prior
Convolutional Decoder ConvDe
Semi-Supervised Learning: Self-Tr
Self-Training for MRI Recon
Why is ConvDecoder Better? Zero Unrolled ConvDecoder
ConvDecoder Noisy Student
Self-Training Takeaway Keep forward model identical; modify pseudo-labels
Invariance to Forward Model Change Supervised Unsupervised
Generalized Consistency Framew
Artifact Invariant Reconstruction
Metrics for Evaluation Discordance between quantitative and qualitative metrics
New Multi-Task Dataset
Surrogates without Dense Labels Use self-supervised tasks to build image representations
Self-Supervised Image Quality
Overall Takeaway • Adding domain knowledges reduces data requirements
Open Source Data/Code
Conclusions
Taught by
Stanford HAI