Liver vessel segmentation is to pinpoint out the composing pixels of vessels from a medical image generated by computed tomography (CT), magnetic resonance imaging(MRI), etc; see the figure alongside. This task has increasingly importance in diagnosis formulating, surgical planning and radiotherapy. However, due to the small size of vessels, it is painful to delineate vessels from livers, hence the scarcity of high-quality and large volume of data.
Here we present LiVS, a fine-grained hepatic vascular dataset constructed by Dr. Zhao’s lab. It has 532 volumes and 15,984 CT slices having vessel masks. The vessels of each slice are delineated by three senior medical imaging experts and the final mask is their majority voting. Due to the pretty small size of vessels, the delineation of each vessel can be oscillating a lot easily. To handle this problem, the coincidence of each vessel among the three masks is calculated. In case the majority voting over any mask is smaller than 0.5, the vessel is highlighted and sent back to the three experts for further refinement. This procedure repeats until no inconsistence exists.
Experiments conducted on this data show that the averaged dice score is 12.8% higher than the existing best. We believe it can be of great help to this area. The details of each volume are described as follows, and the whole data can be downloaded here.
| See | Cases | Sex | Age | StudyDate | SliceThickness(mm) | ImageSize(pixels) | LiverDensity | Spaces | Description |
|---|
