Workshop on Large-scale Annotation of Biomedical data and Expert Label Synthesis
Current storage capabilities and computational power have motivated the rise of data-driven and machine learning algorithms for facilitating the analysis of medical data. However, the acquisition and labelling of training data for biomedical applications presents very specific challenges
. First, labelling often requires trained skills, demanding to keep a medical expert in the loop. Moreover, in numerous applications, the imaged data consists of 3D volumes, which are more time-consuming to annotate than 2D images. Finally, and as a result of the previous, medical experts have to realise a compromise between the size of datasets and the accuracy of their labels, which in return limits the accuracy of predictions of the resulting trained models.
With the challenges above as motivation, this workshop aims at gathering scientific contributions towards:
- The reduction of labelling efforts during the construction of new medical datasets,
- The effective use of existing labelled or unlabelled data from images or other sources,
- An increased accuracy of the provided labels, including fusion of annotations from several sources and better assessment of inter-expert variability and label uncertainty.
- Workshop: October 21, 2016 (PM)
- Full paper submission deadline:
June 20th, 2016 EXTENDED until the 1st of July .
- Notification of acceptance: August 1st, 2016.
- Camera-ready submission: August 10th, 2016.
- Room: Athenaeum CC I
- Location: Intercontinental Athenaeum
Objectives and Scientific Topics
There has been an increasing interest of the MICCAI community for data-driven methods such as supervised learning techniques. The effectiveness of such approaches often depends on their access to sufficiently large quantities of labelled data of good quality. Despite the increasing amount of acquired clinical data, the availability of ready-to-use annotations is very limited. A first purpose of this workshop is to raise awareness on the importance of a methodological acquisition of training data and a careful design of the labelling procedures. A second goal is therefore, to promote the development and scientific exchange of algorithms that focus on assisting the annotation process by making it for example more general, more accurate, faster or more intuitive for the medical experts. To this end, we will include in the program an introductory keynote speech, while calling for paper submissions addressing the labelling/annotation task by means of approaches from the following fields:
Because data annotation and expert labelling is strongly grounded in practical considerations, we welcome not only research contributions, but also encourage submitters to share war stories and practical feedback on successful or insightfully unsuccessful data collection experiences in real world settings
- Active learning
- Semi-supervised learning
- Reinforcement learning
- Domain adaptation and transfer learning
- Crowd-sourcing annotations and fusion of labels from different sources
- Data augmentation
- Modelling of label uncertainty
- Visualization and human-computer interaction
Contact: To reach out to all workshop organizers you may use the LABELS mailing list via