Machine learning techniques have recently received interest as a means of improving MRI reconstruction. Conventionally, ill-conditioned reconstruction problems are solved using iterative optimization algorithms that alternate between applying data consistency and a proximal operator based on a regularizer. This iterative procedure can also be unrolled for a finite number of iterations to generate a feed-forward model. In physics-driven machine learning approaches, the known forward encoding model is used for enforcing data consistency in an unrolled iterative regularized least squares reconstruction. A neural network, which may or may not share weights across different unrolled iterations, is used as the regularizer prior. In this study, we aim to compare several neural network architectures, namely U-Net, ResNet and DenseNet for such physicsdriven reconstruction. The performance of these architectures are evaluated on the publicly available fastMRI knee database. Comparisons are made for uniform and random undersampling masks. The results indicate that a DenseNet regularization unit performs as well as the other strategies for both uniform and random undersampling patterns, even though it has considerably fewer trainable parameters.
|Original language||English (US)|
|Title of host publication||2019 IEEE 10th Annual Information Technology, Electronics and Mobile Communication Conference, IEMCON 2019|
|Editors||Satyajit Chakrabarti, Himadri Nath Saha|
|Publisher||Institute of Electrical and Electronics Engineers Inc.|
|Number of pages||5|
|State||Published - Oct 2019|
|Event||10th IEEE Annual Information Technology, Electronics and Mobile Communication Conference, IEMCON 2019 - Vancouver, Canada|
Duration: Oct 17 2019 → Oct 19 2019
|Name||2019 IEEE 10th Annual Information Technology, Electronics and Mobile Communication Conference, IEMCON 2019|
|Conference||10th IEEE Annual Information Technology, Electronics and Mobile Communication Conference, IEMCON 2019|
|Period||10/17/19 → 10/19/19|
Bibliographical noteFunding Information:
This work was partially supported by NIH, Grant numbers: P41EB015894, P41EB027061; NSF, Grant number: CAREER CCF-1651825. The first two authors contributed equally to this work.
© 2019 IEEE.
- Data consistency
- Deep learning
- MRI reconstruction
- Parallel imaging
- Recurrent neural networks
- Unrolled network