Learn2Reg 2023

We are pleased to host the Learn2Reg Challenge again this year!
In summary, L2R 2023 will include the following tasks:

1. Continuation of the previous tasks NLST , LungCT , OASIS , AbdomenMRCT
2. Registration of lung images of CBCT and FBCT from multiple time points.

With L2R 2023, we focus on two aspects of the actual use of medical image registration methods: practical, clinically-relevant registration tasks and cross-dataset usability.
We address the difficult but highly relevant registration problem of image guided radiation therapy (IGRT) between pre-therapeutic fan beam CT (FBCT) and interventional low dose cone beam CT (CBCT). We provide at least 20 scan pairs with semi-automatic labels for training and manual landmarks for testing. In addition, we will evaluate the submitted algorithms on a private, unpublished real-world dataset from the University Hospital L├╝beck (UKSH) to assess their direct use in clinical practice. We plan to introduce this task as a comprehensible entry point for future type III challenges.

Lung nodule tracking is a practical and highly relevant task in longitudinal lung cancer screening. Therefore, we will extend last year's NLST task by including tracking of nodules as an evaluation criterion (as a TRE).

A principle of Learn2Reg has been to lower the barrier of entry into the field of medical image registration by providing pre-processed data through normalization and pre-registration. However, this leads to specialized algorithms that rely on this specific preprocessing and have limited direct use for other tasks and datasets. In L2R 2023, we are therefore releasing the original image data to allow participants to design their complete registration pipeline. However, we are also making our previous preprocessing available as open source.

In addition to publishing algorithms and methods, the principles of open science also include ease of trial and application. In L2R 2023, we encourage all participants to use Grand Challenge Algorithms to provide their algorithms. GC algorithms provide the ability to deploy methods without publishing their source code and still allow deployment on other people's data.

We plan to publish a set of classical and DL-based baselines as a registration library as GrandChallenge algorithm and use them for our tasks. By providing easy-to-use solutions applicable to a variety of medical registration problems, we hope to strengthen both generalization and comparability between them.
As in the previous years, prize money will be awarded to the best solutions!



NLST (w/ Nodules)


Congratulations to all winners!