Difference between revisions of "2017 Winter Project Week/Needle Segmentation from MRI"

From NAMIC Wiki
Jump to: navigation, search
Line 36: Line 36:
 
<div style="width: 27%; float: left; padding-right: 3%;">
 
<div style="width: 27%; float: left; padding-right: 3%;">
 
<h3>Progress</h3>
 
<h3>Progress</h3>
 
+
* A plenty of useful discussions
 +
* The workflow was modified in order to segment the entire needle instead of just the tip
 +
* Good preliminary results (see picture)
 +
* Start to increase the patch size and the number of negative patches
 +
* Start to test U-net strategy
 
</div>
 
</div>
 
</div>
 
</div>

Revision as of 17:01, 12 January 2017

Home < 2017 Winter Project Week < Needle Segmentation from MRI

Key Investigators

  • Ziyang Wang, BWH
  • Alireza Mehrtash, BWH
  • Alireza Ziaei Torbati, BWH
  • Guillaume Pernelle, Imperial College London
  • Paolo Zaffino, Magna Graecia University of Catanzaro, Italy
  • Andre Mastmeyer, University of Luebeck, Germany
  • Tina Kapur, BWH/HMS

Project Description

This project is a continuation of the project started during the 2016 summer project week ( Needle Segmentation from MRI), NeedleFinder offers tools to segment needles from MRI/CT. It has mostly been tested on MRI from GYN brachytherapy cases. Currently the user must provide the needle tip for the segmentation to start. We aim to detect the tip automatically to make the needle segmentation fully automatic.

Objective

  • Automatic detection of needle tips in MRI from GYN brachytherapy cases.

Approach, Plan

  • We have manually segmented around 1k needles from GYN brachytherapy cases. We want to use this data in a supervised learning approach.
  • We have a pipeline for preprocessing the data (image spacing and normalization)
  • TODO: choose between 2 strategies:
    • Continue with the binary classification approach -> in the region of interest, voxels are classified as positive (needle) or negative (background). Challenge: link those voxels together to form needles / filter out false negatives. Alternative: try to find the tip position and use the previous algorithm to segment the needle
    • Use a semantic segmentation approach (fully convolutional neural networks [J Long 2015], Region-based semantic segmentation with end-to-end training [Caesar 2016])

...

Progress

  • A plenty of useful discussions
  • The workflow was modified in order to segment the entire needle instead of just the tip
  • Good preliminary results (see picture)
  • Start to increase the patch size and the number of negative patches
  • Start to test U-net strategy