Back to Projects List
We plan to prepare a Lung RadioTherapy patient cohort for deep learning segmentation and annotation. We will first use Slicer and other NA-MIC tools to register patients’ follow-up scans to their original planning CT scans as a preparatory step for deep learning. if time allows, we will train a deep neural network to predict patient outcomes from the time series data.
The Planning CTs have excellent annotations:
But since follow-up scans are different times, there is no registration between them. The CTs are “miles apart”. Shown below is the annotations from the planning CT superimposed over a follow-up CT, showing the shift between the different patient scans:
After a prelimary registration in Plastimatch, the anatomy annotations are much closer, as shown below. The image shows the original segmentation objects superimposed over a registered follow up scan taken 3 months later. Because of the time between scans, there was actual morphology changes to the anatomy as well. This result was encouraging after trying only a few parameter exploration attempts. Since Plastimatch operations can be scripted, this approach can automate registration for multiple patient scans in a cohort:
Below is a snapshot of how the segmentation mask for the moving image is growing to match the anatomy and mask in the fixed image. Our deep learning registration results this week don’t match as well as using traditional methods, but this is an emerging application area for deep learning that will continue to improve. Thanks, Neha!
We also learned that giving a registration system incorrect parameters can warp an moving image too much. After generating a strangely warped image by mistake, we just gave it some coloring to create art. Here are our project team’s two submissions to the “Project Week 35 3D-Slicer Art Competition”. Vote for your favorite. Vote by editing this page or vote on Curt’s facebook page…
Votes for #1: 0
Votes for #2: 0
The Image Data Commons has datasets with annotations across multiple time points, so this is an available dataset to practice registration techniques. Free Google Cloud credentials are available for experimenting without having to download data for processing. Simply select the cohort through IDC for analysis: https://imaging.datacommons.cancer.gov/explore/?filters_for_load=%5B%7B%22filters%22:%5B%7B%22id%22:%22120%22,%22values%22:%5B%22qin_prostate_repeatability%22%5D%7D%5D%7D%5D
Registration tools have been added to Project-MONAI in the 0.5 release. The MONAI tutorials include a registration example now, which we used as a basis for our experimentation: