Deidentifying MRI Data Domain by Iterative Backpropagation.

Medical images acquired at various hospitals can differ significantly in their data distribution. We can find multiple sources of divergence evaluating images from different clinical centers, patient diseases, vendors, or even configurations on the same scanner. Typically at deployment, when we are facing real-world domain, data is collected from it and the trained model is adapted. This is not practical in all scenarios like medical imaging due to the lack of data and the strict protection to which it is subjected. We investigate this challenging problem by evaluating a novel domain adaptation procedure. First, a classifier model is trained to distinguish between which data distribution comes from. Once trained, the images from each vendor are modified iteratively using the gradients of the error obtained when the target is set arbitrarily. Finally, when we have a new sample we only have to carry out the same process of domain adaptation by error backpropagation. The experiments were performed on Multi-Centre, Multi-Vendor & Multi-Disease Cardiac Image Segmentation Challenge (M&Ms), comparing the segmentation metrics obtained for studies from vendors present in the training set and an additional studies from an unseen vendor.