Automatic CADS, TotalSegmentator and AtlasNet Segmentations of the Porcine Hepatic CT Dataset (Kainz et al., ISBI 2014)

Word of caution. These masks are unreviewed predictions from models trained on human CT, applied here to pigs. They are not ground truth, and no accuracy figures (Dice or similar) exist. Pig anatomy differs from human anatomy, so a mask with a plausible shape can still be mislabelled. Use them as first drafts only, and check them before any use. Not for clinical or diagnostic use. Automatic segmentations of all 171 CT scans (57 pigs, 3 contrast phases) of the porcine hepatic CT dataset by Kainz et al., ISBI 2014 (doi:10.1109/ISBI.2014.6867814). The CT images themselves are in a separate record under their own licence and are not included here. Contents Pig_CADS_segmentations.tar: CADS 1.0.3 "research" weights, tasks 551-559 (1539 label maps, 9 per scan), with label map, volume tables, overlays and scripts. Pig_vessel_masks_all.tar: 342 masks, TotalSegmentator 2.18.0 liver_vessels and AtlasNet organs for every scan, with comparison pages, volume tables and scripts. montage_all_pigs_pv.png, montage_vessels_pv.png, vessel_comparison_all_pigs.pdf: overview figures. The PDF has one page per pig (portal-venous scan), CADS against both other models. README_CADS_segmentations.txt, README_vessel_masks.txt: the READMEs from inside the two archives. Orientation 23 pigs (69 scans) were stored upside-down, and their NIfTI headers do not record this. They were head-foot flipped for inference and the masks were flipped back, so every mask aligns with the original file. Left-right is supported by the predicted anatomy, not by acquisition records. The list of flipped pigs is in flipped_pigs.txt in each archive. Any earlier copy of the CADS package has mirrored labels on those 69 scans and must be replaced. Known problems Pig26, Pig31 and Pig97_pv have suspect header spacing, so their volumes are unreliable. CADS can label the left part of the pig liver as spleen. Pig23 and Pig27 stay low in some phases, and Pig44 has no CADS aorta. The CADS portal/splenic vein label is fragmentary. TotalSegmentator liver_vessels also labels the IVC and does not separate portal from hepatic veins. AtlasNet found no celiac trunk in any scan. Human label names do not map one-to-one onto pig anatomy; vertebra, rib and lung lobe names are unreliable. See the READMEs for the full list and the method. Citation Please cite this record, the dataset paper, and the model you use: CADS (Xu et al., arXiv:2507.22953), TotalSegmentator, or AtlasNet. Licence All files in this record, including the overview figures, are released under CC BY-NC-SA 4.0. This replaces the "no licence chosen" and "check before release" notes in the READMEs. It is a cautious choice, because it is unsettled whether a model's licence carries over to its output masks. The models used are licensed as follows: CADS research weights CC BY-NC-SA 4.0, AtlasNet CC BY-NC 4.0, TotalSegmentator liver_vessels Apache-2.0. The CT images remain under CC BY 4.0 in their own record.