Difference between revisions of "2017 Winter Project Week/Population Based Image Imputation"

From NAMIC Wiki
Jump to: navigation, search
 
(3 intermediate revisions by the same user not shown)
Line 7: Line 7:
  
 
==Key Investigators==
 
==Key Investigators==
*Adrian Dalca, MIT
+
*Adrian Dalca, MIT, MGH
 
*Katie Bouman, MIT
 
*Katie Bouman, MIT
 +
*Mert Sabuncu, MGH, Cornell
 
*Polina Golland, MIT
 
*Polina Golland, MIT
  
Line 28: Line 29:
 
* Implement algorithm updates on GPU  
 
* Implement algorithm updates on GPU  
 
* investigate banding side-effects.
 
* investigate banding side-effects.
 +
|
 
<!-- Progress and Next steps bullet points (fill out at the end of project week) -->
 
<!-- Progress and Next steps bullet points (fill out at the end of project week) -->
*
+
* Implemented entire algorithm on GPU.
 +
* Comparing 4-CPU with a GTX1080, we gain significant speedup when dimensionality of subspace is large, but in the realistic scenario (subspace has low dimensionality) the speedup is only ~2x.
 +
* Banding in images found to be due to bad cluster assignment of patches. We're working on assigning cluster assignment of patches based on original subject space rather than atlas space.
 
|}
 
|}
  
 
==Background and References==
 
==Background and References==
 
<!-- Use this space for information that may help people better understand your project, like links to papers, source code, or data -->
 
<!-- Use this space for information that may help people better understand your project, like links to papers, source code, or data -->

Latest revision as of 14:36, 13 January 2017

Home < 2017 Winter Project Week < Population Based Image Imputation

Key Investigators

  • Adrian Dalca, MIT, MGH
  • Katie Bouman, MIT
  • Mert Sabuncu, MGH, Cornell
  • Polina Golland, MIT

Project Description

We developed a model for image imputation - or restoration - for clinical quality images where slice separation (e.g. 6mm) is significantly larger than slice resolution (e.g. 1mm^2). Our model captures statistical correlations within a collection of clinical images from a population of subjects at each location in the image. This means we learn different model parameters for many image locations involving methematical updates that involve many small matrix multiplications. In this project we want to investigate the potential for GPUs to help in the runtime of the algorithm.

Objective Approach and Plan Progress and Next Steps
  • To test (and implement) a GPU version of Population Based Image Imputation, which learns a sparse-observation low-dimensional Gaussian Mixture Model at many locations in an image.
  • Test whether the type of operations we do (most many many small matrix multiplications and look-up) can be made significantly more efficient on a GPU
  • Implement algorithm updates on GPU
  • investigate banding side-effects.
  • Implemented entire algorithm on GPU.
  • Comparing 4-CPU with a GTX1080, we gain significant speedup when dimensionality of subspace is large, but in the realistic scenario (subspace has low dimensionality) the speedup is only ~2x.
  • Banding in images found to be due to bad cluster assignment of patches. We're working on assigning cluster assignment of patches based on original subject space rather than atlas space.

Background and References