Curriculum-learning for Vessel Occlusion Detection in Multi-site Brain CT Angiographies

Andrés Martínez Mora, Michael Baumgartner, Gianluca Brugnara, Maximilian Zenk, Yannick Kirchhoff, Aditya Rastogi, Alexander Radbruch, Martin Bendszus, Clara I. Sánchez, Philipp Vollmuth, Klaus Maier-Hein

Keywords: Curriculum Learning, Domain Shift, Medical Object Detection

Abstract: Deep learning models often fail to generalize to target data due to shifts between target and training data distributions. Hence, their impact in the real world is limited. To solve this, including training data from more sites may not be enough, as new data may also present shifts with the original training data, complicating the learning process. We hypothesize that curriculum-learning may provide more robust models against training site shifts by sorting these sites in order of increased difficulty. In this work, we focus on Vessel Occlusion detection in CT angiographies from stroke-suspected patients from three sites, training first on large homogeneous balanced sites, which we hypothesize are easier to learn. Next, we incorporate small heterogeneous imbalanced sites, which may be more complex. Our approach is compared to training only on a large homogeneous site (single-site training) and to training on all sites (pooled-site training). We reach a 2% improvement in FROC and AUROC scores. Thus, adequately ordering the training sites based on simple characteristics such as label balance or data size may improve model robustness.

Submission Number: 109