This research presents a new annotation strategy, Drag&Drop, for tumor detection and localization. This strategy simplifies the annotation process making it more efficient, particularly for temporal and volumetric imaging. The research proposes that weak annotations are more appropriate for tumor detection and localization than per-pixel annotations, as they are cost-effective and require less time. The experimental results show that the Drag&Drop method achieves better detection and localization performance than alternative weak annotations. Furthermore, with limited resources, allocating weak annotations from a diverse patient population fosters models more robust to unseen images than allocating per-pixel annotations for a small set of images.
Publication date: 23 Oct 2023
Project Page: https://github.com/johnson111788/Drag-Drop
Paper: https://arxiv.org/pdf/2310.15098
