2016 Winter Project Week/Projects/ImageRestoration

• 

  Projects List

• 

  Clinical Stroke Image

• 

  "true" high res data
  

• 

  More results show our ability to improve images, but only on small areas.

Key Investigators

• Adrian Dalca (MIT)
• Katie Bouman (MIT)
• Polina Golland (MIT)

Project Description

Most synthesis, in-painting or super-resolution methods require a training dataset which includes the desired-quality images. Unfortunately, in the clinical setting this is often now available.

Due to the low quality of clinical images (often with many artifacts, 7mm thick slices, etc), most standard algorithms, such as those for registration, segmentation, analysis, will fail.

To improve results for large datasets of clinical-quality data, we are investigating restoration methods without training datasets. Here, we are using a patch-based Gaussian Mixture Model approach with MRF priors and utilizing only the current dataset, without an external training dataset.

Essentially, we explore a model where for a given location in all volumes of a dataset, we model those image patches as drawn from a particular mixture model.

Objective	Approach and Plan	Progress and Next Steps
We will investigate a current model for learning and using Patch-based Gaussian Mixture Model to restore sparse-slice data.	We've implemented a small version of the model without regularization and that only works on a small section of a volume. We'll try to deploy it on large-scale dataset on entire volumes.	We've built infrastructure for deploying the training at each location in parallele We've used our lab's cluster to deploy training with various parameter at each of 18,000 locations We've coded reconstruction modules to rebuild each region, but still need to combine each of the reconstructions.