ESTRO 2024 - Abstract Book

S4424

Physics - Machine learning models and clinical applications

ESTRO 2024

TRAQinform IQ was able to identify ROIs for targeted radiation therapy that would increase the average predicted PFS by at least 100 days as determined by TRAQinform Profile. Further investigation is warranted to estimate the true clinical impact of this method.

Keywords: Outcome prediction, targeted ablation

663

Digital Poster

Reproducibility guideline for deep learning research in medical imaging

Attila Simkó 1 , Anders Garpebring 1 , Joakim Jonsson 1 , Tufve Nyholm 1 , Tommy Löfstedt 2

1 Umeå University, Department of Radiation Sciences, Umeå, Sweden. 2 Umeå University, Department of Computing Sciences, Umeå, Sweden

Purpose/Objective:

Concerns about the reproducibility of deep learning research are more prominent than ever, with no clear solution in sight. For medical imaging applications, prestigious venues expect state-of-the-art solutions [1], high quality papers and appreciate publicly available supplementary material, however the quality of the latter receives little to no attention. This creates a large drawback of the proposed solutions with respect to their usefulness in further connected research or their clinical evaluation for applications in radiotherapy. Peer-reviewed papers from the research field often offer online, publicly available material which suggests transparency and that the solutions are readily available for other researchers. We question this assumption, assess frequently encountered issues in public code repositories and offer guidance for researchers.

Material/Methods:

Between 2018 and 2023, the Medical Imaging with Deep Learning (MIDL) conference has peer-reviewed and accepted 428 full-length research papers in the fields of medical imaging. These papers explore automated solutions, for e.g. guiding radiotherapy treatment through segmentations, synthetic CT generation, etc. We have carried out a systematic evaluation of all papers exploring their data availability, and evaluating if all essential components of their published code is available or not. As a baseline, we have used already established resources for reproducibility [2, 3, 4], expanding and adjusting them to the commonly encountered issues during the evaluation of the papers. Due to the large interest in our findings, we have presented an online workshop where the guidelines were discussed, improved, and a poll was conducted allowing us to learn more about the severity of our concerns.

Results: