Frequently Asked Questions
How do I register? ▼
Please follow the instructions here.
Which citations should i use for the pretraining dataset? ▼
Please follow the Citation Policy.
Where can i find the Synapse website? ▼
The Synapse website can be found here.
Computational Resource Requirements ▼
We have tried our utmost to ensure that the computational resources required for the challenge are minimal. Users are expected to have access to one or more GPUs with at least 40GB of memory as well as a cluster with at least 1.5TB of storage. Given access to one H100 GPU, pretraining using the provided code takes approximately 48 hours depending on your I/O performance. Finetuning takes approximately 10 hours. Further, the codebase is setup such that it is easy to extend to multiple GPUs and multiple nodes using DDP, where we have seen linear scaling up to 8 GPUs even on multiple nodes.
Ethical Approvals ▼
All datasets have been approved by legal and ethics boards, including the Danish Patient Safety (Styrelsen for Patientsikkerhed, approval #31-1521-257) and the Danish Data Protection Agency (Datatilsynet, approval #P-2020-320). The pre-training source dataset denoted as "MGH Wild" was downloaded from the PACS at Massachusetts General Hospital under Institutional Review Board approval.
How do i submit my models and what is required to submit? ▼
Submission will be through Synapse. Instructions will be provided when we open the submission portal. Final
submission will require a short writeup including a description of the methodology (track 1 and 2) and data
used (track 2). The writeup should be no more than 2 pages long.
Submission will be using Apptainer containers. To ensure your container adheres to the requirements, you can
use
this container
validator. More info will come soon.
How will you process models? ▼
Models will be processed on a secure platform physically located in Copenhagen which is ISO-27001 certified, the highest atainable certification for information security and personal data protection. Models will only be used for purposes of the challenge and subsequently deleted. The challenge organizers are able to sign confidentiality agreements when necesarry. Please contact us for more info.
Am i allowed to do supervised pretraining in the open track? ▼
Yes, you are allowed to do supervised pretraining in the open track. However, there are two important constraints: First, the same pretrained model must be finetuned to each of the three downstream tasks (as illustrated in figure 1). Second, no additional data is allowed during the finetuning phase -- you must use only the provided finetuning datasets.
Am i required to make my code publicly available? ▼
It is not a requirement to publish your code, however it is a requirement to be invited to the follow-up paper that the method is described in necessary detail to reproduce the results. During submission, we will have a questionnaire which will ask detailed questions about the submission. Please note, that we of course highly encourage all participants to publish their code.