Uploaded July 2025 | Updated November 2026, 1 hour ago
Recent machine learning-based reconstruction methods exploit paired (e.g., corresponding regular-dose and low-dose CT scans) or unpaired datasets to learn data models or reconstruction mappings. Incorporating additional priors into the learning framework could enable models learned with limited training data to generalize better.
In this talk, we will first review an approach for model-based image reconstruction (MBIR), where the sparsity-based regularizer's parameters are learned in an unsupervised manner from a small set of high-quality images. We pre-learn a union of sparsifying transforms that can cluster image patches into multiple groups, with a specific transform well-matched to each group. When incorporated in the regularizer in MBIR, the learned transforms provide much better image reconstructions in CT compared to conventional filtered backprojection and non-adaptive regularization methods especially at low X-ray doses.
We then extend the approach to a unified supervised-unsupervised (SUPER) learning-based scheme that combines classical MBIR optimization and unsupervised transform learning regularization together with supervised deep learning in a common formulation. We provide multiple interpretations of the resulting scheme from fixed point iteration analysis or bilevel training optimization and show that with limited training data, it provides much better CT image reconstructions at low X-ray doses than the constituent supervised or unsupervised schemes. Finally, we present a combined supervised-unsupervised framework involving dictionary-based blind or on-the-fly learning and deep supervised learning for MR image reconstruction from under-sampled k-space data. The results show that the image features captured by the blind dictionary-based approach complement deep network learned features to provide significantly better reconstructions with limited training data.
Video created 12/06/2021
Recent machine learning-based reconstruction methods exploit paired (e.g., corresponding regular-dose and low-dose CT scans) or unpaired datasets to learn data models or reconstruction mappings. Incorporating additional priors into the learning framework could enable models learned with limited training data to generalize better.
In this talk, we will first review an approach for model-based image reconstruction (MBIR), where the sparsity-based regularizer's parameters are learned in an unsupervised manner from a small set of high-quality images. We pre-learn a union of sparsifying transforms that can cluster image patches into multiple groups, with a specific transform well-matched to each group. When incorporated in the regularizer in MBIR, the learned transforms provide much better image reconstructions in CT compared to conventional filtered backprojection and non-adaptive regularization methods especially at low X-ray doses.
We then extend the approach to a unified supervised-unsupervised (SUPER) learning-based scheme that combines classical MBIR optimization and unsupervised transform learning regularization together with supervised deep learning in a common formulation. We provide multiple interpretations of the resulting scheme from fixed point iteration analysis or bilevel training optimization and show that with limited training data, it provides much better CT image reconstructions at low X-ray doses than the constituent supervised or unsupervised schemes. Finally, we present a combined supervised-unsupervised framework involving dictionary-based blind or on-the-fly learning and deep supervised learning for MR image reconstruction from under-sampled k-space data. The results show that the image features captured by the blind dictionary-based approach complement deep network learned features to provide significantly better reconstructions with limited training data.
Video created 12/06/2021










