guest ::
| Home > Publications database > LesionLocator: Zero-Shot Universal Tumor Segmentation and Tracking in 3D Whole-Body Imaging |
| Home > Publications database > LesionLocator: Zero-Shot Universal Tumor Segmentation and Tracking in 3D Whole-Body Imaging |
| Preprint | DKFZ-2025-00592 |
; ; ; ; ; ; ; ; ;
2025
arXiv
Abstract: Abstract: This archive contains the Lesion Dataset with Synthetic Follow-ups, which provides original and synthetic second-timepoint images with annotations as part of the LesionLocator paper, a framework for zero-shot universal tumor segmentation and tracking in 3D whole-body imaging. The dataset is approximately 700 GB in size and contains around 5,200 images. It is designed to support research in lesion tracking, segmentation, and progression analysis. Lesions are annotated with instance-based labels, ensuring consistent lesion identification across both timepoints. The dataset includes images sourced from multiple publicly available datasets, covering a variety of lesion types and anatomical regions. Due to constraints related to image size, quality, or licensing, not all images were included in the final dataset. This dataset is particularly well suited for pretraining or for use in combination with real longitudinal imaging data to improve model generalization. For longitudinal tracking tasks, we recommend introducing image misalignment through data augmentation, such as cropping one timepoint, to better simulate real-world conditions. A more detailed description of the dataset is available in the paper or the archive, downloadable via the "Fulltext" link at the bottom of the page.
Keyword(s): Computer Vision and Pattern Recognition (cs.CV) ; Artificial Intelligence (cs.AI) ; FOS: Computer and information sciences
|
The record appears in these collections: |
