Improved image segmentation of bone by image filtering

Project Description

Can we do a better job of segmenting the trabecular structure in knee scans by HR-pQCT…

The use of HR-pQCT of the knee is relatively new, and although we are able to collect images as part of our normal protocol now, the size of the knee seems to lead to some beam hardening issues at the centre. This is manifested as a slight darking of the trabecular structure. There is work by the UCSF group (Kazakia, Burghardt) that has shown the so-called Laplace-Hamming filter provided by Scanco Medical IPL seems to perform better than the standard Gaussian filtering. Other methods such as Otsu’s method, adaptive filtering, etc, are also options.

The goal of this project is to compare filtering methods on the ability to extract bone microarchitecture. We would want to determine how these affect measures such as TbTh and TbSp, and possibly finite element outcomes. This could be explored using the TRIKNEE study where we have three repeat measures on the same individual for a small cohort. The ground-truth would be taken as the Gauss segmentation, since we don’t have a true gold standard for these in vivo scans.

We have implemented a L-H filter here: https://github.com/Bonelab/Bonelab/blob/master/bonelab%2Fcli%2Ffft_laplace_hamming.py

Scope

• PhD/MSc side project
• Summer student

Data Source

• TRIKNEE

Resources

• IPL scipts
• Python scipts