LUS-7Seg: A Multi-Annotation Dataset and Benchmark for Standard Plane Classification and Anatomical Segmentation in Hepatobiliary Ultrasound
Description
LUS-7Seg is a multi-annotation dataset designed for multi-class standard plane classification and anatomical segmentation in hepatobiliary ultrasound. It contains 10,507 high-quality 2D images from 2,415 healthy subjects, covering 7 core hepatobiliary standard planes. Crucially, each image is equipped with expert-verified, pixel-level masks for the liver parenchyma and gallbladder. This dataset is intended to advance ultrasound computer-aided diagnosis from coarse image-level classification to interpretable, structure-aware localization.
Files
Steps to reproduce
Detailed preprocessing pipelines, training configurations, and evaluation scripts for the benchmark baselines (ResNet, ViT, U-Net, nnU-Net) are provided in the associated open-source repository. Please refer to the "Benchmarking Methodology" section of the main manuscript for comprehensive experimental setups.
Institutions
- University of Technology SydneyNew South Wales, Sydney