Student project update: Segmentation error

Posted in: Student projects

Here at the University of Bath, our semester has come to an end (it's gone by so fast!), so it seems like the perfect time to update you on the student project work that has been going on. In particular I want to focus on an exciting project that one of our Integrated Mechanical and Electrical Engineering students, Abdallah Mahmoud, has been working on.

Most biomechanics finite element models start with some sort of imaging (such as a CT scan or an MRI scan) which can be segmented to provide the patient-specific three-dimensional geometry of the different tissues. Specialised software is used to perform the segmentation (such as Mimics from Materialise, or ScanIP from Simpleware). In theory this part of the process could be completely automated, as algorithms have been developed which make use of machine learning techniques, but these are currently limited to certain anatomical sites and come with additional cost. For these reasons manual segmentation is still very common. The purpose of this project was to quantify how much variation and error was introduced during the segmentation process.

In order to look at error, it was necessary to perform the segmentation multiple times - so Abdallah spent quite a long time at the computer segmenting the geometries. He looked at three different ankle scans, and segmented each three times, and then comparisons were made with models segmented by a user very familiar with ankle anatomy to provide a 'gold standard' measure. Various metrics (such as bone volume and cartilage thickness) were then used to quantify the accuracy and the reliability of the segmentation.

Bone regions highlighted in red in a CT scan illustrating the segmentation process
Segmentation of ankle bones from a CT scan

The results showed Abdallah's overall reliability was excellent (ICC>0.9), despite being a novice user of the software. However, there was a big reduction in reliability when it came to measurements of the short distances, such as distances between the talus-navicular and the talus-tibia. We believe that this reflects the importance of using high quality resolution of ankle scans to ensure reliable and accurate geometries between the bones. From a finite element perspective this would be critical for studies exploring contact stresses or cartilage damage. So the good news is that the software is user-friendly enough for an inexperienced user to generate reliable and accurate geometries, and the bad news is that a useful FE model cannot be created from a clinical CT/MRI scan with poor resolution if you want good data.

If you want to hear more about this study, please get in touch (email We are looking into disseminating the findings at an upcoming conference and will post the details if the abstract is successful!

Posted in: Student projects


  • (we won't publish this)

Write a response