Where is VALDO - Vascular Lesions Detection Challenge 2021

Clinical Motivation and methodological context

Appropriate blood supply is essential to the healthy maintenance of brain tissue. With age, vascular changes are observed in the smallest vessels resulting in impaired function. Changes to the surrounding tissue can be observed using magnetic resonance imaging. White matter hyperintensities (WMH) are one such prominent marker of cerebral small vessel disease (CSVD) and their automated segmentation has been the focus of a large body of research as well as of segmentation challenges. Other markers of CSVD exist and their quantification along with WMH is essential to grasp the overall picture of the vascular burden related to CSVD. They include notably lacunes, enlarged perivascular spaces and cerebral microbleeds. Manual annotations are extremely time-consuming and suffer greatly from inter- and intra-rater variability, due to their small size and the difficulty of distinguishing these markers from each other and similarly appearing structures as well as the lack of a way to uncover the "real" ground truth. However, many studies have hinted at their potential to become essential biomarkers. Automated methods are therefore required to make their quantification not only robust and reliable, but simply feasible. So far development of such methods has been impeded by the methodological issues related to their very small size and the extreme imbalance in the data, but also the absence of sufficient gold standard.

This challenge aims at promoting the development of new solutions for the automated segmentation of such very sparse and small objects while leveraging weak and noisy labels. The central objective of this challenge is to facilitate quantification of CSVD in brain MRI scans.

The challenge will have a technical impact in the following fields: object detection, segmentation, class imbalance, use of weak labels, multi-scale object detection, assessment of prediction uncertainty. The biomedical impact will not only directly impact the field of cerebral small vessel disease research but also other brain pathologies such as multiple sclerosis where similar objects have recently been shown renewed interest. More broadly, translation of developed techniques to other fields where sparse object detection is essential will be impacted.


Note: The challenge is separated into 3 different tasks designed to tackle key issues related to the detection, segmentation and characterisation of small markers of CSVD:

            Task 1: Enlarged PVS                                                            Task 2: Microbleeds                                                                 Task 3: Lacunes

Task #1 Enlarged perivascular space (PVS) segmentation and counting

The burden of enlarged PVS is currently emerging as an important neuroimaging biomarker. The current bottleneck for studying PVS burden is the need for an automated method. Manual annotation of enlarged PVS is extremely time-consuming due to the large number of enlarged PVS that can be present in MRI scans. Currently, neurological studies mostly score the burden of enlarged PVS visually by e.g. counting the number of enlarged PVS in a slice. This is the most practical and fast way to quantify enlarged PVS, however it is a coarse approximation of the large amount of valuable information in the scans. Furthermore, manual annotation and visual scoring are subject to observer bias due to the difficulty of assessing if a PVS is enlarged and distinguishing it from other similarly appearing structures. Dealing with this subjectivity of annotations is one of the main challenges for current automated methods, as it is not possible to acquire a "real" ground truth.

A robust, automated method for segmenting enlarged PVS would be extremely useful for neurological research on the role of enlarged PVS in neurological disorders.

 Cerebral microbleed detection and segmentation
Cerebral microbleeds are an essential marker of cerebral small vessel disease. Their presence has been associated with specific vascular pathology such as cerebral amyloid angiopathy and with other markers of cerebral small vessel disease (WMH, enlarged PVS). Currently, microbleeds are largely identified manually. The challenges of automated identification of microbleeds are the presence of numerous mimics and the sparsity of the data: microbleeds are very small and most often there are very few microbleeds per scan.

An automated method for microbleed segmentation would enable further research on their presence in the context of neurodegenerative diseases.

 Lacunes detection, segmentation and uncertainty
Lacunes of presumed vascular origin are another important biomarker for cerebral small vessel disease. Currently, lacunes are generally detected manually, which is very time-consuming and subjective. Important challenges for automated methods are the small size of lacunes and their rare occurrence. Lacunes can look very similar to enlarged PVS; they are often mistaken for each other. Even for experts distinguishing lacunes from enlarged PVS can be challenging and sometimes impossible depending on the MRI scan quality. This is more problematic for lacunes than for enlarged PVS, as the prevalence of lacunes is substantially lower than the prevalence of enlarged PVS. Automated methods should additionally focus on modeling the detection and segmentation uncertainty, as this is especially important for this biomarker.

An automated method for lacune segmentation would facilitate further research on lacunes and their role in neurological diseases.

Where is VALDO was accepted for the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2021. The challenge results will be presented during a half-day satellite event at MICCAI 2021. After this we will show the results in public leaderboards on this website. The challenge design document can be found here (important heads up: this document has recently been updated). 


For any information or queries contact Carole Sudre or Kimberlin van Wijnen at valdo.challenge.2021@gmail.com