The NeoBrainS12 challenge
Goal of the NeoBrainS12
Neonatal brain tissue volumes are considered to be an indicator of long-term neurodevelopmental performance. Accurate brain tissue segmentation is a prerequisite for obtaining volume measurements.
The goal of NeoBrainS12 is to compare the performance of (semi-)automatic algorithms for neonatal brain tissue segmentation in T1- and T2-weighted MRI scans.
Thus far, a number of different algorithms have been presented but comparing their performance reliably is not feasible. The algorithms were evaluated on different sets of images, with differences concerning both the acquisition protocol and the composition of the patient population, and the algorithms were evaluated by different criteria. Since obtaining a reference standard for brain segmentations has proven extremely time-consuming and cumbersome, very few methods have been evaluated on complete scans; the performance of most methods was assessed on a (very) limited subset of scan sections. In addition, the published algorithms did not consistently segment the same tissue (sub-)types; for example, some algorithms distinguish between cortical and central grey matter, whereas others lump these structures into one segment. Reliable re-implementation of various algorithms for the purpose of comparison would be very difficult, if at all feasible.
In this study all methods are compared against the same reference standard. This provides an opportunity to compare the performance of different segmentation algorithms.
For the segmentation T1-weighted and T2-weighted MR images of the brain acquired with a 3T MRI scanner are provided. Three different image sets of pre-term born infants are available:
- axial scans acquired at 40 weeks corrected age
- coronal scans acquired at 30 weeks corrected age
- coronal scans acquired 40 weeks corrected age
Given T1- and T2-weighted sequences, the task is to segment
- cortical grey matter
- basal ganglia and thalami
- unmyelinated white matter
- myelinated white matter
- cerebrospinal fluid in the extracerebral space
(Semi-)automatic segmentations are evaluated against a manually set reference standard in the complete scans.
Each team is free to decide which data set and how many tissue types to segment. Each tissue type needs to be segmented according to the definition provided here. The results are published on this web page. In addition, we ask you to submit a paper describing your method by mailing it to email@example.com. Submitted papers are limited to maximum 8 pages (suggested length: 4 pages). Please let us know if the paper has been submitted for review elsewhere and therefore cannot appear online.
Please include the following in the submitted papers:
- Is your algorithm automatic or semi-automatic? If user input is needed, please describe this.
- Give the overall structure of the algorithm and describe each step of the algorithm. If pre- or post-processing is used, please describe that.
- List limitations of the algorithm. Is the algorithm specifically designed to segment only certain types of scans? E.g. is the algorithm designed for segmenting healthy or pathological images, for term-born or pre-term infants scanned at certain gestational age, scanned using a specific protocol?
- Was the algorithm trained with example data? If so, describe the characteristics of the training data.
- If the algorithm has been tested on other databases, you could consider including those results.
- What is the average runtime of your algorithm, and on which system is this runtime achieved?