Advancing the frontiers of deep learning for low-dose 3D cone-beam CT reconstruction

As part of ICASSP-2024 Signal Processing Grand Challenges, we are hosting a 3D cone-beam computed tomography (CBCT) reconstruction challenge using deep learning-based methods. The challenge seeks to push the limits of deep learning algorithms for 3D cone beam computed tomography (CBCT) reconstruction from low-dose projection data (sinogram). The key objective in medical CT imaging is to reduce the X-ray dose while maintaining image fidelity for accurate and reliable clinical diagnosis. In recent years, deep learning has been shown to be a powerful tool for performing tomographic image reconstruction, leading to images of higher quality than those obtained using the classical solely model-based variational approaches. Notwithstanding their impressive empirical success, the best-performing deep learning methods for CT (e.g., algorithm unrolling techniques such as learned primal-dual) are not scalable to real-world CBCT clinical data. Moreover, the academic literature on deep learning for CT generally reports the image recovery performance on the 2D reconstruction problem (on a slice-by-slice basis) as a proof-of-concept. Therefore, in order to have a fair assessment of the applicability of these methods for real-world 3D clinical CBCT, it is imperative to set a benchmark on an appropriately curated medical dataset. The main goal of the challenge is to encourage deep learning practitioners and clinical experts to develop novel deep learning methodologies (or test existing ones) for clinical low-dose 3D CBCT imaging with different dose levels.

The challenge is now live! Submission deadline is 10 December, stay tuned for submission instructions.

Registration:

Please register for the challenge to get a link to the dataset. You can access the information about the data in the "Data" tab on this webpage.

Evaluation Criteria:

Two winners will be selected, one for the low-dose data and one for the clinical dose data. We encourage you to submit to both. We will rank by the average MSE on the test set (unreleased). Your reconstructed images should be of the same size as the provided samples, i.e. 256^3 volumes.

Winners will be invited to present their work in a special session at ICASSP-2024, in Seoul, Korea.

FAQs:

Q: Are only ML methods accepted?

No, any methods, including classical approaches, will be accepted.

Q: When is the deadline?

10 December 2023, so short time! Get the code working soon!

Q: Is there any required format to submit the models/code?

Yes, we will be releasing some conda environments for ML inference. If they don't fit your purposes, please contact us and we will try to accommodate reasonable alternatives (as long as we can install them ourselves locally).

Q: Will there be a live leaderboard?

Unfortunately no. Due to the high cost of inference (110 CBCT images), maintaining a live leaderboard could get out of hand quickly without significant resources.

For any further queries, please feel free to contact us.