The LLSD dataset was constructed to evaluate cross-dataset generalization in laparoscopic liver landmark segmentation. It was derived from the Laparoscopic Liver Resection (LLR) dataset [1], which contains 46 frames of real surgical procedures from 4 patients. Each frame was annotated with multi-class segmentation masks, where each pixel is assigned to one of three landmark classes (anterior ridge, silhouette, and falciform ligament) or background. Compared with L3D [2], LLSD was annotated with thinner, centerline-like landmark masks in order to emphasize boundary localization. This difference in annotation protocol results in slightly lower overlap scores (e.g., DSC, IoU) when models trained on L3D are evaluated on LLSD.CITATION AND REFERENCES[1] Rabbani, N. et al. (2021) ‘A methodology and clinical dataset with ground-truth to evaluate registration accuracy quantitatively in computer-assisted laparoscopic liver resection’, Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 10(4), pp. 441–450. doi:10.1080/21681163.2021.1997642.[2] Pei, J. et al. (2024) ‘Depth-driven geometric prompt learning for laparoscopic liver landmark detection’, Lecture Notes in Computer Science, pp. 154–164. doi:10.1007/978-3-031-72089-5_15.