Learn2Reg - The Challenge (2023)

Medical image registration plays a very important role in improving clinical workflows, computer-assisted interventions and diagnosis, and research studies involving e.g. morphological analysis. Besides ongoing research into new concepts for optimisation, similarity metrics, domain adaptation and deformation models, deep learning for medical registration is currently starting to show promising advances that could improve the robustness, generalisation, computational speed and accuracy of conventional algorithms to enable better practical translation. Nevertheless, before Learn2Reg there was no commonly used benchmark dataset to compare stateof-the-art learning-based registration among another and with their conventional (not trained) counterparts. Since last year, the Learn2Reg challenge is a type 2 challenge meaning that only submissions containing the algorithm (e.g. Dockers) are allowed. This facilitates reproducibility and further use of the algorithms in the research community.

The most prevailing unmet requirements for clinical adoption of medical image registration are two-fold: 1) limited generalisation of one registration model to an unseen task, 2) limited clinical focus of evaluation metrics. We strongly believe that the extensions we propose for our challenge design in 2023, can tackle both aspects and further bridge the gap between research and practical use.

We could demonstrate initial successes in the area of developing self-configuring registration algorithms that can automatically be trained and optimised on a variety of unseen datasets. This step is crucial, since challenge datasets can never comprehensively mirror all aspects of a certain clinical application and for registration to
become a universal tool the hyper-parameter choices should become part of an automatic pipeline. Several methods were submitted that produced robustly good results on the three hidden datasets. However, a number of algorithms - especially deep-learning-based ones - still had problems working equally well on all datasets.

In terms of evaluation metrics, our previous designs had either limited clinical relevance (e.g. anatomical landmarks in lung scans are only a proxy for relevant tumour lesions) or limited coverage, i.e. using either target registration error or Dice overlap exclusively cannot necessarily detect deterioration in quality outside the supervised structures.

For this reason, we extend the tasks from 2022 in certain aspects of Learn2Reg 2023 with a particular focus on expanding the evaluation metrics, adding more hidden datasets and simplifying the setup of self-configuring registration algorithms.

Innovations in Task 1 and 3.
In Task 1, the challenge is divided into two phases: In phase 1, participants train/tune their algorithms locally and submit the algorithms via grand-challenge. The best teams of this phase are invited to participate in phase 2. In phase 2, the participants submit a training docker that will be run by the organizers on a larger dataset that includes additional annotations that are not publically available. The trained networks will be made available via grand-challenge. Specifically the expanded datasets for lung / throrax registration will comprise: anatomical landmarks, geometric keypoint correspondences, therapeutically relevant target structures and semantic anatomical segmentations.

We improve Task 3 from L2R 2022 by introducing changes to facilitate implementing and testing a selfconfiguration registration docker. Innovations include a set of baseline algorithms to guide implementation, an additional dummy dataset with supervision and evaluation configuration files to fully comprehend our evaluation method, as well as predicted semantic label information at test time. We further introduce measures to facilitate a successful submission, including the possibility of sanity checks and a improved timeline for implementation and refinement.