MICCAI Hackathon

Task 1 - Domain shifts between clinical annotators

Medical imaging datasets are known to be highly heterogeneous, being often collected at multiple institutions and resulting in so-called domain-shifts between institutions. At the level of manual segmentations, annotators trained in one institution may follow slightly different annotation procedures, e.g. over- and under-segmenting, and this may bias the learning process. In this task, you investigate the domain shifts in white matter (WM) lesion segmentation.

• Discuss, identify and quantify the impact of multiple sources of shift within the multiple sclerosis dataset of WM lesions in Shifts 2.0 publicly available dataset.
• Quantify the impact of multiple annotators as opposed to a single annotator in terms of model uncertainty. How does model uncertainty vary against the number of annotations considered to create the “golden reference”? Does the uncertainty decrease if more annotators are used?
• How many annotators in the given dataset over-segment the “golden reference”? How many undersegment? What is the observed bias introduced by each of the annotators?

Task 2 - Unsupervised target structure localization

Researchers often face limited or no training data when translating medical image computing methods to a new clinical application. In this task, you aim at localizing target structures in head CT to facilitate manual annotation and provide a region of interest for existing methods working on regions of interest.

• Propose an approach that can automatically localize the target structures in head CT by only having 1 case with manual annotation plus 10 unannotated cases.
• Is the use of pre-trained models feasible? Are traditional approaches outperforming DL-based techniques? Can one benefit from the anatomical dependence among the structures?

Task 3 - Integrated diagnostics

Images are not the only sources of information available to physicians when treating patients. Similarly, in the domain of medical AI, “integrated diagnostics” hypothesizes that combining complementary diagnostic data might yield more robust/better performing predictive models is termed. In this task, you investigate integrated diagnostics.

• Discuss and summarise potential pitfalls in simple integration strategies like logits averaging, or feature vector concatenation.
• Discuss the potential of more sophisticated methods to address this problem and develop a proof-of-concept for mapping different sources of heterogeneous data to a patient outcome.
• Motivate your proposed solution in case of i) missing data (e.g., the results of one of the diagnostic exams is missing) and ii) longitudinal data (e.g., some of the diagnostic exams are repeated over time to check if the treatment is working).

Task 4 - Associations between clinical semantics and deep features

The comprehension of the way deep learning models achieve certain performances remains a barrier to applying this powerful tool in the clinical environment. In this task, you perform an associative analysis of features extracted by pre-trained neural networks and semantic features annotated by clinicians in CT lung nodules, aiming to infer possible relationships linking these feature spaces.

• Identify possible ways to search within the unstructured (deep) feature space to find relationships.
• Propose different feature analysis methods for dependencies evaluation.
• Investigate how sensitive are the results considering different architecture choices.
• Discuss the results, showing if there are semantic features more expressively "found", and what could explain that.

Task 5 - Time-series quality control of medical device manufacturing

Plastic injection molding is a technology for manufacturing single-use medical devices. Generally, statistical process control is solely used after every manufacturing batch to guarantee product quality. However, a continuous control during a manufacturing batch is desired to reduce potential scrap. In this task, you aim at building a mapping between a time-series of manufacturing parameters and product quality.

• Analyze sources of variance (e.g., properties of plastic granules) that have a direct or indirect influence on the product quality and assess.
• Develop a proof-of-concept for mapping manufacturing parameters to product quality.

Bonus task - Do It Yourself

Do you have your own task in mind? Then do it yourself (DIY)! Simply work on your own data/problem or use one of our suggested datasets. Team up with others who might be interested in exploring a dataset and some hacking. Note that you will not have dedicated mentors for support available.

You could work on one of the datasets freely available at The Cancer Imaging Archive (TCIA). Some TCIA datasets that come with data beyond medical images, and, therefore, might be closer to a clinical setting, can be found in the resources.

Participation

• Who can participate? Everyone. The more diverse the better. So, it does not matter whether you are a clinician, a researcher, an undergraduate student, a postdoc, or a professor.
• What do I do as participant? You will work on a task you choose from a set of tasks during the hackathon weekend. Mentors will provide feedback and guidance to you in dedicated mentoring sessions. You will shortly present your contribution at the end of the MICCAI Hackathon.
• Can I participate individually? Yes, working individually on a task is possible, although you would miss the chance to meet new people.
• Can we participate as a team? Yes. To do so, each member needs to fill in the same team name in the registration form.

Registration

• Why do I need to register for the hackathon? A registration is required to facilitate the communication between the organizers and the participants.
• Is a hackathon registration binding? No, you can unregister until the start of the hackathon weekend on September 3.
• Until when can I register for the hackathon? Registration for the MICCAI Hackathon will be open until the hackathon weekend on the September 3 at 09:30 UTC.
• Can I update my registration? Yes, you can always change your registration such as joining a team. Your registration can be updated through the link in the welcome e-mail. Keep in mind that this might affect other people (e.g., switching a team)!
• Is it mandatory to attend the meeting at the satellite event to participate? No, you can participate at the hackathon weekend without attending the meeting at the satellite event day.
• Does the hackathon weekend imply additional fees? No, the participation in the MICCAI Hackathon is free of charge. However, as the MICCAI Hackathon is an official satellite event at the MICCAI conference, attending the meeting at the half-day satellite event underlies the registration fees of the satellite event day.

Organization

• Can I work on my own task? Yes, as long as it is related to the topic of the hackathon.
• Can I work on multiple tasks? Yes, you can combine the ideas of several tasks to form a new task.
• Can I start working on the task before the hackathon weekend? Sure, as soon as you have your team (or decide to work alone) you can start hacking.
• Can I be matched into a team? No, finding your team is up to you. You can however contact the organizers to put you in contact with existing teams that you might join.
• When will the data be available? All relevant material, including a recording of the session, will be made publicly available a week before the hackathon weekend.
• Do I need to choose a mentor? No, mentors get assigned to teams based on the task.

Format

• What platform is used for the hackathon? The platform for the virtual hackathon weekend will be Discord with Zoom for presentations. The meeting at the satellite event on will happen on the official platform or in person.
• What does the award consist of? The award consists of Amazon vouchers sponsered by the Center for Artificial Intelligence in Medicine (CAIM), University of Bern. The first prize contribution will receive a voucher valued 500 USD. In case of eight or more contributions, a second (300 USD) and third (200 USD) place prize will also be awarded.
• Who gets awarded? Independent judges will grade the contributions during the presentations of the contributions. The grading involves three aspects: creativity, applicability, and estimated effort; each graded from 1 (worst) to 5 (best). The best contribution will win an award.
• How do the presentations look like? The presentations consist of a 5 minute recording of your outcome in which you should promote your idea/finding/solution.
• Will everything be publicly available? We encourage participants to open source their outcomes. Final presentations, keynotes, and all provided materials will be available.
• The timezone is suboptimal for me, what can I do? You are free to work on your task whenever you like, but try to benefit from the interactions with your mentor and your teammates.

General

• What happens with all the contributions at the end of the hackathon? The best contributions are invited to present during the meeting at the satellite event day.
• Who has the rights on the outcomes? The outcome is free to use (open source) and participants as well as mentors might decide to follow up on their ideas.
• Is the MICCAI Hackathon a regularly occurring event? Yes. It's already the third edition and we aim at having it recurring. Contact us if you have ideas for the next edition or are interested to help organizing.