Workshop on 3D Segmentation in the Clinic:
- A Grand Challenge -
Date:
|
29. October ' 07 |
Location:
|
Brisbane, Australia
|
| News |
| Dates |
| Program |
| Data |
| Evaluation |
| Organizers |
| Download |
| Proceedings |
Final number of received letters of intent:
Caudate: 26
Liver: 35
Format of submission for segmentation results:
In order to compare all segmentation methods at an equal level, whether they use voxel-level segmentations, or sub-voxel surface based segmentations, all methods need to submit a voxel level segmentation. This is also due to the fact that the manual segmentations, which are used as a quasi gold-standard to compare against, are on a voxel-level. Thus, you have to submit a binary volume of the same size and spacing as the corresponding gray-level data, in which the background is labeled as zero (all other values are treated as object). The file has to be stored in RAW format (e.g. MHD from ITK) with 8 bits per voxel. A value of zero is interpreted as background, all other values are treated as object. In order to automatize the analysis on our end, we need you to store one file for each segmentation (i.e. the left and right caudate structure need to be saved in different files). In case you are using a surface-based segmentation method, you could use ITK's TriangleMeshToBinaryImageFilter for conversion to a volume.
Caudate data - General comments
All MRI images are scanned with an Inversion Recovery Prepped Spoiled Grass sequence on a variety of scanners (GE, Siemens, Phillips, mostly 1.5 Tesla). Some datasets have been acquired in axial direction, whereas others in coronal direction. All datasets have been re-oriented to axial RAI-orientation, but have not been aligned in any fashion. All data is stored in Meta format containing an ASCII readable header and a separate raw image data file. This format is ITK compatible. Full documentation is available here.
Caudate training data
The training data is from 2 major sources:
S-1) MRIs and structural segmentations from the internet brain segmentation repository (ibsr) at Mass General Hospital, Boston
S-2) MRIs and caudate segmentations from the Psychiatry Neuroimaging Laboratory at the Brigham and Womens Hospital Boston.
The definition of the caudate in these 2 datasets is not exactly the same, so that there are slight differences between the caudate definitions, especially in regard to the border with the nucleaus accumbens anteriorly and the vanishing tail posteriorly. You can choose to develop a model based on both training populations (and incorporate the variability between the 2 manual segmentation protocols), or only to train on one of the models. Please state clearly in your paper what training population your model is built on.
contains dataset from IBSR repository
located at http://222.cma.mgh.harvard.edu/ibsrcontains the original IBSR datasets
Left caudate label: 11
Right caudate label: 50Original Readme from IBSR distribution contains the 18 reoriented IBSR datasets
Left caudate label: 1
Right caudate label: 2contains 14 datasets from BWH PNL group
grayscale images and segmentations afterLeft caudate label: 1
Right caudate label: 2Caudate testing data
There will be 2 sets of testing data, one distributed before the workshop and one after the workshop. The datasets are gathered together from several sources:
S-2) 14 MRIs from the Psychiatry Neuroimaging Laboratory at the Brigham and Womens Hospital, Boston. This data is from the same study as the S-2 datasets in the training set
S-3) MRIs from a Parkinsons Disease study at the UNC Neuro Image Analysis Laboratory, Chapel Hill
S-4) MRIs from a pediatric study at the UNC Neuro Image Analysis Laboratory, Chapel Hill
S-5) MRIs from a test/re-test study at the UNC Neuro Image Analysis Laboratory, Chapel Hill and the Duke Image Analysis Laboratory, DIAL
The diversity of the test data will test your algorithms in respect to a variety of properties:
a) flexibility wrt to pathology, age group, and signal-to-noise aspects,
b) stability in a test/re-test situation.
c) stability vs differences in caudate definition as the manual segmentations of these datasets is based on slight differences in the protocol, especially in regard to the border with the nucleaus accumbens anteriorly and the vanishing tail posteriorly.
Liver data - General comments
All CT images are enhanced with contrast agent and scanned in the central venous phase on a variety of scanners (different manufacturers, 4, 16 and 64 detector rows). As it is CT, all datasets have been acquired in transversal direction. The pixel spacing varies between 0.55 and 0.8mm, the inter-slice distance varies from 1 to 3mm. There is no overlap between neighboring slices. All data is stored in Meta format containing an ASCII readable header and a separate raw image data file. This format is ITK compatible. Full documentation is available here.
Liver training and testing data
A total of 40 images is used for the workshop. Most of them are pathologic and include tumors, metastasis and cysts in different sizes. Images have been randomized into three groups: 20 training images, 10 testing images for the qualifying and 10 for the contest. The downloadable archive consists of the training images (including reference segmentations) and the first 10 testing images (without segmentations).
All segmentations were created manually by radiological experts, working slice-by-slice in transversal view. The first tool they employed was an intensity-based region grower. In case of leakage, these leaks were removed by drawing manual cut-lines. The segmentation is defined as the entire liver tissue including all internal structures like vessel systems, tumors etc. In general, a vessel counts as internal if it is completely surrounded by liver tissue (in the transversal view). The large vessels that enter the liver (V.Cava and portal vein) are segmented in the part which is enclosed by liver tissue, i.e. as the convex hull of the liver shape in that area.