2016 Winter Project Week/Projects/PatchRegistration

• 

  Projects List

• 

  Clinical Stroke Image

• 

  Current Resulting registration on isotropic Buckner images illustrated via segmentation overlap from an atlas.

Key Investigators

• Adrian Dalca (MIT)
• Andreea Bobu (MIT)
• Polina Golland (MIT)

Project Description

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, and analysis, will fail. In part, this is because registration algorithm depend on assumptions of smooth anatomical structures and good quality images, which are not present in these sparse clinical acquisitions. Here, we are investigating a patch-based discrete image registration which allows for more versatile image metrics and does not impose similar assumptions.

Objective	Approach and Plan	Progress and Next Steps
We will investigate a current implementation for patch-based discrete registration on sparse-slice data.	We have an implementation of patch-based discrete registration using a standard patch distance metric. We'll start by evaluating this current metric on isotropic images We'll develop the infrastracture and evaluate the current metric on isotropic data at large. We'll develop a new metric for sparse images and test it out.	We set up registration on a buckner40 dataset with a large number of parameters and launched on a large cluster We found a global setting that yields very good results, with volume overlap (DICE) of most subcortical structures in the 80s and 90s. Results shown above