Energy-efficient Medical Image Processing - E2MIP 2024

To curtail and reduce the impact that climate change has on our socio-economic live, saving energy is key. Data centers in general and modern-day AI applications in particular are electricity super-users. Recent studies have attempted at estimating the carbon footprint of common large-scale AI applications, highlighting the unsustainable, environmentally questionable path of current AI research. Despite this research, reducing or even monitoring the energy consumption needed for computational approaches in medical imaging are still poorly investigated. We counteracted to this situation, i.e. the model development purely under the perspective of predictive performance while disregarding the accompanied environmental consequences, by the first round of the Energy-Efficient Medical Image Processing challenge during the last MICCAI. Given the ongoing importance of this topic, we aim to continue this important line of research. Especially given the focus of this years MICCAI on developing countries and the huge financial burden of high computing/ high energy solutions.

The goal of the challenge is to raise awareness for energy consumption of training and inference methods and foster the development of novel best-practice approaches and solutions to improve the energy efficiency of commonly used DL/ML/MIP models. This will hopefully increase the awareness of energy consumption in needed for medical image processing and lead to novel approaches that allow more efficient algorithms. In addition to this, we try to gather more information about the current situation w.r.t. energy-efficient computation on medical image processing, for example, the ratio of training and inference runs with an additional survey. Towards this end, the challenge will offer two pathways to develop energy-efficient medical image processing models:

• The original challenge will call for the submission of training and inference on a dedicated public dataset (the actual training/test-split is hidden to the participants) for three common tasks: segmentation, detection, and classification.
• To foster best practices and reporting of energy consumption in general AI model development, co-submission for inference and possibly training for other challenges will be offered. This was also done last year as a post-challenge analysis for a few challenges and we aim to build on the already established collaborations.

Each submission will be evaluated on the tier-2 Supercomputer HoreKa, located at the Karlsruhe Institute of Technology (KIT), Germany. HoreKa allows for precise measurements of whole compute node energy consumption per run via internal power sensors that are part of Lenovo's XClarity Controller (XCC) and can be read via IPMI. The whole workload energy consumption of submitted solutions will be measured for a full training run and inference on the hold-out test set. Submissions will be offered a dedicated amount of computing resources (1 full node, equipped with 4 NVIDIA A100 GPUS with each 48 GB of VRAM, connected via NVLINK, and Intel Xeon Platinum 8368 CPUs with a total of 76 cores) to run training and testing. For energy measurements, the participants have to submit their solutions before the on-site event. The final solution will then be calculated and evaluated in a period of two to three weeks before MICCAI and the final results reported during the on-site event. To allow the participants a prior estimation of the success of their approach, we will present guidelines and tools to measure the energy consumption on standard hardware during an envisioned initial workshop and on the challenge website. 

We expect a trade-off between required energy consumption and achieved performance. To account for this, we will not report a winning approach but rather a Pareto front between achieved performance and energy consumption. A selection of high-performance approaches on the Pareto front will be given the chance to present their solution and approaches during the on-site event. The actual selection will depend on the number of submissions; however, the presentations will be selected w.r.t. interest for the audience, performance, completeness of pre-experiments, and originality.