Self-Supervised Learning for 3D Medical Imaging (SSL3D) Challenge

his challenge seeks to identify the most effective self-supervised learning (SSL) method for 3D medical imaging. By standardizing critical factors such as pre-training dataset, network architecture, preprocessing pipelines, and fine-tuning schedules, the competition ensures a fair and transparent evaluation environment that allows to identify the best SSL pretext task for 3D medical imaging.

The development of SSL techniques for 3D medical imaging has been hindered by inconsistencies in research practices. Variability in pre-training dataset sizes, differences in network designs, and the use of diverse downstream evaluation datasets have created significant barriers to identifying the most effective methods. This lack of standardization complicates the comparison of approaches and limits progress in the field. This challenge seeks to overcome these obstacles by offering a common pre-training dataset and fine-tuning framework to fairly evaluate self-supervised pre-training strategies and drive innovation in 3D medical imaging.

Dataset Overview
Participants will have access to the largest publicly available head & neck MRI dataset derived from the publicly available OpenNeuro platform. This dataset is an amalgamation of over 700 smaller datasets, all published under a permissive CC0 or PDDL license. After cleaning and standardization of image format and metadata, the final dataset comprises 114,570 3D volumes from 34,191 patients. The collection contains a variety of well-established MRI sequences, such as T1w, Flair, T2w sequences, as well as diffusion coefficient (ADC) and fractional anisotropy (FA) maps. More details about the dataset can be found in the pre-print associated with this challenge https://arxiv.org/pdf/2412.17041.

Task Description
Participants can utilize two predefined network architectures: one based on convolutional neural networks (CNNs) and the other on transformer models. The objective is to develop SSL methods that enable these networks to learn robust and generalizable feature representations. Submissions will consist solely of the pre-trained weights, which will be assessed through fine-tuning on at least 6 hidden head & neck MRI segmentation and classification tasks. These evaluation tasks will remain confidential to ensure an unbiased and standardized comparison of methods. The challenge participants will not have access to the private downstream datasets and the finetuning will be done by the challenge organizers.