Image reconstruction by domain-transform manifold learning

• Bo Zhu  |  Bruce R. Rosen  |  Jeremiah Z. Liu  |  Matthew S. Rosen

• Harvard Medical School Boston USA

Magnetic resonance imaging (MRI), X-ray computed tomography (CT), positron emission tomography (PET), ultrasound, and radio astronomy all rely on image reconstruction.Since analytic knowledge of the exact inverse transform does not exist a priori, image reconstruction can be difficult, particularly in the presence of sensor non-idealities and noise.AUtomated TransfOrm by Manifold APproximation (AUTOMAP) is an image reconstruction model that recasts image reconstruction as a data-driven, supervised learning task that enables a mapping between the sensor and image domain to emerge from a suitable corpus of training data.AUTOMAP is implemented with a deep neural network and exhibit its flexibility in learning reconstruction transforms for a variety of MRI acquisition strategies, using the same network architecture and hyperparameters.The training dataset of generic images was assembled from ImageNet18. 10,000 images from the “Animal,” “Plant,” and “Scene” categories along with that dataset of brain images was assembled from the MGH-USC Human Connectome Project (HCP)17 public database.The AUTOMAP uses MRI as a model system in this demonstration, but this model is applicable to image reconstruction problems in a wide variety of modalities.AUTOMAP provides a powerful new paradigm for image reconstruction implemented with a deep neural network that learns an optimal reconstruction function for any acquisition strategy and exhibits superior performance for noisy acquisitions compared with conventional approaches.

Input variables : Image from MRI, X-ray, PET, ultrasound
Output Variables : Reconstructed Image

Visit Model : github.com

Additional links : nature.com