Ground Truth for Segmentation

Some of the challenges in creating the ground truth for segmentation have to do with the fact that it is rarely routinely created in clinical practice. Radiation oncology is one of the few fields where segmentation is generated as part of the treatment path, but normally segmentation projects require custom labeling efforts.

One of the things to keep in mind when dealing with a labeled (segmented) dataset is that interpretation of radiological images is ambiguous and quite often, two independent clinicians (observers) would not label things in the same way. This phenomenon is called Interobserver Variability and has been studied in the literature.

• This is the paper that I have mentioned where the authors present results of measuring the variability between radiation oncologists segmenting structures in the head and neck region: Mukesh, M et al. “Interobserver variation in clinical target volume and organs at risk segmentation in post-parotidectomy radiotherapy: can segmentation protocols help?.” The British journal of radiology vol. 85,1016 (2012): e530-6. doi:10.1259/bjr/66693547
• While I worked on Microsoft’s Project InnerEye we also did our own IOV study for how conformant people are in contouring pelvic anatomy in prostate cancer patients, and included it into our paper which you can read here: Macomber, M. W., Phillips, M., Tarapov, I., Jena, R., Nori, A., Carter, D., … Nyflot, M. J. (2018). Autosegmentation of prostate anatomy for radiation treatment planning using deep decision forests of radiomic features. Physics in Medicine & Biology, 63(23), 235002. doi: 10.1088/1361-6560/aaeaa4

When it comes to tooling for creating ground truth, 3D Slicer is a popular free tool used in the research community, and I will walk you through using it for creation and review of segmentation labels in the next lessons. MITK is another one. However, many medical imaging startups and larger companies use tools of their own.