RepConTUR AbdomenAtlas Annotations

Official annotations for "RepConTUR: A Report-Conditioned Volumetric Multimodal Model for Abdominal Tumor Segmentation".

This dataset extends AbdomenAtlas 3.0 Mini to suborgan-level promptable tumor segmentation and report-derived visual question answering. It holds annotations only: the CT volumes and the original masks must be downloaded from AbdomenAtlas 3.0 Mini.

• Code: https://github.com/Oncosoft-Research/RepConTUR
• Model: https://huggingface.co/Oncosoft-Research/RepConTUR

Contents

annotations.json covers the 3,798 AbdomenAtlas 3.0 Mini CT volumes that contain a liver, kidney or pancreas tumor.

	Count
CT volumes	3,798 (3,425 train, 373 test)
Tumor masks	13,469 (4,843 whole-organ, 8,626 suborgan)
Referring phrases	10 per mask
VQA pairs	71,783 (48,451 yes/no, 13,433 short answer, 9,899 multiple choice)

Each mask targets one location:

Organ	Whole organ	Suborgan
Liver	all liver	hepatic segments 1–8 (Couinaud)
Kidney	all kidney	the left kidney, the right kidney
Pancreas	all pancreatic	the pancreatic head, body, tail

The four tasks of the paper use these annotations as follows:

Task	Input	Target
Referring segmentation (RES)	CT + a referring phrase	the tumor mask
Report-grounded segmentation (GSEG)	CT + narrative report + a referring phrase	the tumor mask
Report generation (RRG)	CT	the narrative report
Visual question answering (VQA)	CT + a question	the answer

Format

annotations.json follows a COCO-like layout:

info          {description, min_lesion_size}
categories    [{id: 1, name: liver_lesion}, {id: 2, name: kidney_lesion}, {id: 3, name: pancreatic_lesion}]
images[]      id, bdmap_id, split (train | test), file_name (image_only/<bdmap_id>/ct.nii.gz),
              demographics {sex, age, contrast}, reports {structured_report, narrative_report},
              vqa[] {id, question, answer, type, options (multiple choice only)}
annotations[] id, image_id, category_id, location, size_voxels,
              source_mask (mask_only/<bdmap_id>/segmentations/<organ>_lesion.nii.gz),
              rle_per_slice[] (one COCO RLE per axial slice), sentences[] {sentence}

All paths are relative to the AbdomenAtlas 3.0 Mini root. contrast is null when the phase is not recorded. Each mask is stored as one COCO RLE per axial slice. It decodes to a (W, H, D) binary volume on the voxel grid of the volume's ct.nii.gz, in the same orientation as the NIfTI data array:

import json
import nibabel as nib
from pycocotools import mask as mask_utils

dataset = json.load(open('annotations.json'))
images = {image['id']: image for image in dataset['images']}

annotation = dataset['annotations'][0]
image = images[annotation['image_id']]
tumor_mask = mask_utils.decode(annotation['rle_per_slice'])            # (W, H, D) uint8
ct = nib.load(f'AbdomenAtlas3.0Mini/{image["file_name"]}').get_fdata()  # same shape
print(annotation['location'], annotation['sentences'][0]['sentence'])

Use with the RepConTUR code

Place the AbdomenAtlas 3.0 Mini volumes and masks under data/abdomen_atlas/{image_only,mask_only} in the code repository, then:

hf download Oncosoft-Research/RepConTUR annotations.json --repo-type dataset --local-dir data/abdomen_atlas

The training configs read data/abdomen_atlas/annotations.json. The repository also contains the scripts that rebuild this file from AbdomenAtlas 3.0 Mini (suborgan_assignment, then vqa_generation).

How it was built

• Suborgan masks. Each whole-organ tumor mask is split into connected components, and each component is scored against the AbdomenAtlas suborgan masks (liver Couinaud segments, left or right kidney, pancreatic head, body or tail) by intersection over component (IoC) and surface distance. Each suborgan the structured report names takes as many components as the report counts there, best-ranked first. Every remaining component goes to its highest-IoC suborgan and to any other suborgan it overlaps by IoC ≥ 0.2, or to the nearest suborgan if it overlaps none. A component can therefore belong to two adjacent suborgans. The components in one suborgan are merged into one mask.
• Referring phrases. Ten phrasings per mask are drawn from templates for the organ and location.
• VQA. Question–answer pairs are generated from the structured and narrative reports by a language model and checked by a second language model.
• Reports. The structured and narrative reports are those of AbdomenAtlas 3.0. They are synthetic, produced by the RadGPT pipeline, not written by radiologists.

Limitations

• The reports are synthetic. They often give the largest tumor and the tumor count rather than every tumor's location.
• The released file is regenerated with the current scripts and differs slightly from the data the released model was trained on: that run had 8,354 suborgan masks, and its RRG and VQA tasks also used training volumes without a tumor.
• Research use only. Not for clinical decisions.

License

The annotations adapt AbdomenAtlas 3.0 Mini (reports and tumor masks) and are released under the same CC BY-NC-SA 4.0 license: attribution, non-commercial use only, and adaptations under the same license. Cite AbdomenAtlas 3.0 and RadGPT when you use this dataset, as well as:

@article{lee2026repcontur,
  title  = {RepConTUR: A Report-Conditioned Volumetric Multimodal Model for Abdominal Tumor Segmentation},
  author = {Lee, Seunghwan and Park, Sangjoon and Lee, Jongmin and Kim, Yoon Jo and Cho, Wonyoung and Kim, Jin Sung and Chae, Han Joo},
  year   = {2026}
}