MRI Prediction Service
Magnetic Resonance Imaging reconstruction and contrast prediction
Magnetic Resonance Imaging (MRI) is applied in material sciences for non-invasive investigation of sample structure and composition, by leveraging the differences in tissue contrasts. However, every different type of contrast encoded in the MR image typically requires a separate measurement, which is a time-consuming task.
To reduce the number of measurements, the MRI Prediction Service was designed. The service allows users to upload a DICOM file obtained at a given contrast expressed in terms of the pulse sequence parameters TE and TR (echo time and repetition time, respectively), and predicts how the image would look like if measured with an alternative contrast, based on the given theoretical sequence parameters that would have been used.
MRI Prediction Service input page
The resulting image is then displayed as in a preview and can be downloaded together with the corresponding theoretical sequence parameters used to generate it and the original DICOM file. In the case where one DICOM file contains multiple images, an alternative image is predicted for each, but only the first image in the stack is displayed to prevent possible conflicts with large datasets.
MRI prediction result
The service optimizes the information contained within the datasets measured in the same MRI experiment. By predicting an alternative contrast type from a given one, the measurement time can be decreased by a factor n for each of the n contrast types.
Currently, the machine learning model has been trained on predicting contrasts for a solution of CuSO4 at different concentrations, as required by our target users. We are open to receive further input data from the community to train the model on different materials to further improve the applicability of the service.
Contact
Rossella Aversa, rossella.aversa@kit.edu
