Back in 2008, I (along with colleagues at Siemens Corporate Research) invented a system to find and segment tumors in full-body MRI scans. It’s challenging to find all types of tumors across the entire body, but the ability to automatically detect tumors wherever they are can aid early detection and save lives.
We patented the findings in 2008 and received confirmation today that the patent has been granted (#8,155,405). Read on for a quick overview of the approach and a few useful links if you’re interested in seeing how it works!
System and Method For Lesion Segmentation In Whole Body MRI.
Gozde Unal, Gregory G. Slabaugh, Tong Fang, Shawn Lankton, Valer Canda, Stefan Thesen, and Shuping Qing. US Patent Number: 8,155,405. Filed March 2008. Granted April 2012.
The algorithm was constructed with 3 major parts:
- Detecting potential tumors
- Initializing with a “best guess”
- Segmenting with active contours
- Checking validity
My contribution was primarily around initialization and segmentation. Detection and validity checking are performed with bayesian algorithms that can determine likely tumor regions and pass those regions on to the subsequent parts of the system.
The first “best guess” is created using the Grow-Cut algorithm (described here) which is a fast and robust way to find the approximate shape of a detected tumor in 3D.
Next, we applied a tailored formulation of localized active contours (described here, and published on extensively) to refine the segmentation and provide an accurate representation of the location and shape of the tumor.
Many thanks to Gozde, Greg, and Tong for their guidance and help in getting this done!