Ischemic Stroke Lesion Segmentation
Welcome to Ischemic Stroke Lesion Segmentation (ISLES) 2017, a medical image segmentation challenge at the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2017 (10-14th September). ISLES will be held jointly with the Stroke Workshop on Imaging and Treatment CHallenges (SWITCH).
This year ISLES 2017 asks for methods that allow the prediction of lesion outcome based on acute magnetic resonance imaging (MRI) data. Therefore, a multi-spectral data set of 51 stroke patients and matching expert segmentations, manually outlined of follow-up scans, are provided. Please find more details about the challenge's motivation, data, rules and information about how to participate in the paragraphs below.
We encourage participants to attend both stroke-related satellite events at MICCAI 2017. Come in the morning to SWITCH, where technological and clinical aspects are discussed, and subsequently to the ISLES challenge. This is a perfect opportunity to get both technical and clinical insights into current problems concerning stroke.
April | Registration opens and distribution of training data |
1st May | Early bird rate for MICCAI 2017 registration |
1st August, 23:59 GMT | Deadline for submitting abstract |
15th August | Distribution of test data |
29st August, 23:59 GMT | Deadline for submitting test data results |
10th September | Presentation of method and result on the workshop day |
This challenge for stroke lesions segmentation has been very popular the past two years (2015, 2016) and yielded various methods, that help to tackle important challenges of modern stroke imaging analysis. This year the challenge provides acute stroke imaging scans and manually outlined lesions on follow-up scans.
If you are interested in participating, you are invited to download the training set, including both MRI scans as well as the corresponding expert segmentations of stroke lesions. This will allow you to validate and optimise your method as much as you favour.
Shortly before MICCAI 2017 will take place, a set of test cases will be released of which participants will be asked to run their algorithm on and upload their segmentation results in form of binary image maps. To complete a successful participation, participants will need to submit an abstract, describing the employed method.
The organizers will then evaluate each case and establish a ranking of the participating teams. All results will be presented during SWITCH at MICCAI 2017 and will be discussed with invited experts and all workshop attendees.
Each team will have the opportunity to present their submitted method as a poster, while selected teams will be asked to give a brief presentation detailing their approach. Eventually, submissions will be included in the workshops LNCS post-proceedings and potentially compiled for a high-impact journal paper to summarise and present the findings.
Defining location and extend of a stroke lesions is an essential step towards stroke assessment. Of special interest is the change of lesion over time, as this could provide valuable information about tissue outcome after stroke onset. Although, modern magnetic resonance imaging techniques (diffusion/perfusion imaging) can be useful to distinguish between acutely infarcted tissue ("core") and hypo-perfused lesion tissue ("penumbra"), automated methods to do so are sparsely used or too simple to capture the full complexity of the data set. Therefore, there is a great need for advanced data analysis techniques that could help to define these regions and tissue outcome in a more reproducible and accurate way and eventually support clinicians in their decision-making process (e.g. deciding for or against thrombolytic therapy).
Medical image processing comprises many tasks, for which new methods are regularly proposed. However, varying data set size and heterogeneity make it nearly impossible to compare different approaches in a fair way. By providing a high-quality data set publicly and pre-define evaluation rules, challenges like ISLES aim to overcome these limitations and create a common framework for adequate comparison of results.
Images will be available in uncompressed Neuroimaging Informatics Technology Initiative (NIfTI) format: *.nii. All MRI sequences have been skull stripped, anonymized and co-registered for each subject individually. No further pre-processing took place, allowing participants to apply their optimised processing pipelines.
Training data set consists of 43 patients. Developed techniques will be evaluated by means of a testing set including 32 stroke cases. Acquired MRI sequences are described in detail below.
Under-perfused brain tissue can be recognised as hyper-intense regions of the DWI trace images (DWI maps). Contrary the apparent diffusion coefficient (ADC) maps show these regions as dark areas. In comparison, ADC maps do not suffer from confounding T2 shine-through effects as observed on DWI maps.
To assess cerebral perfusion a contrast agent (CA) is administered to the patient and its temporal change is captured in dynamic susceptibility scans. Subsequently, perfusion maps are derived from these raw data for clinical interpretation of perfusion of blood within the brain tissue. Different maps aim to yield different information of perfusion, most commonly calculated maps are: The cerebral blood volume (CBV), cerebral blood flow (CBF) and mean transit time (MTT), defined as the ratio of volume to flow of cerebral blood (CBV/CBF). Furthermore, the time to peak concentration of the CA (TTP) and the time need at which the (deconvolved) residue function reaches its maximum value (Tmax).
For ISLES 2017, those maps were computed by means of Olea Sphere's block decomposition.
To assess the final lesion outcome, an anatomical sequence (T2w or FLAIR) was acquired when the stroke lesion had stabilised. Provided ground-truth segmentation maps were manually drawn on those scans.
By registering, each team agrees to use the provided data only in the scope of the workshop and neither pass it on to a third party nor use it for other publications. After the workshop took place, the data will be released under a research license. No copyright transfer of any kind will take place, except in the case of a contribution to the LNCS post-proceedings special issue.
For data access and result submission, please register at SICAS Medical Image Repository:
Further clinical data for training cases is also available.
Each team wishing to participate in the ISLES challenge is required to:
Any automatic method that predicts the lesion outcome is of interest. There is no restriction on new, innovative or unpublished methods. Semi-automatic methods are eligible for participation and will appear on ranking board, but will not be part of the competition, as it is impossible to rate the influence of the manual steps in a fair manner.
ISLES requires the participating teams to submit an abstract of one page in LNCS format, which will be reviewed by the organizers. The text will be distributed among the MICCAI 2017 attendees on the MICCAI 2017 USB drive and uploaded to the challenge result web-page. See dates for the associated deadlines.
Requirements
Review
Submission
Data distribution, registration and automatic evaluation will be handled by the SICAS Medical Image Repository:
There you will find explanations on how to register, how to download the data and how as well as in which format to upload your results. Furthermore, the evaluation scores obtained by each team will be listed there.
The ranking as presented on the 10th September 2017 at the challenge event is as follows:
The evaluation comprises two evaluation measures, each of them highlighting different aspects of the segmentation quality. From the computed evaluation scores, the global ranking of each team is established as described on ISLES 2015 for the SISS task (without ASSD).
1Computed excluding the failed cases (where Dice=0), therefore not suitable for a direct comparison of methods.
This challenge is organized by
Dr. Arsany Hakim, Department of Diagnostic and Interventional Radiology, Inselspital Bern, Switzerland.
Dr. Mauricio Reyes, Institute for Surgical Technology & Biomechanics, Universität Bern, Switzerland
Prof. Roland Wiest, Department of Diagnostic and Interventional Radiology, Inselspital Bern, Switzerland.
Stefan Winzeck, Division of Anaesthesia, Department of Medicine, University of Cambridge, UK
Christian Lucas, Institute of Medical Informatics, Universität zu Lübeck, Germany