Machine Learning for Image Reconstruction
Making sense of sparse data
The acquisition of raw data for medical images is often time consuming or harmful to patients. Algorithms and techniques that allow us to reconstruct images from datasets that are under-sampled or corrupted by artifacts can enable an improved patient experience and reduce costs. At Microsoft we are engaged in the development of machine learning models that facilitate accelerated imaging. We have enhanced the InnerEye (opens in new tab) toolkit to support training of reconstruction models. Specifically, we are providing dedicated tooling to enable the use of the FastMRI (opens in new tab) dataset and existing model architectures. This includes tooling for managing the large MRI raw datasets needed for training the models. We have also published clinical annotations (opens in new tab) that enable evaluation of new models in the context of pathology.