Datasets:
File size: 3,713 Bytes
06f54eb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | ---
language:
- vi
task_categories:
- visual-question-answering
- image-to-text
tags:
- radiology
- mri
- brain
- medical-imaging
- nifti
size_categories:
- n<1K
configs:
- config_name: mri_reports
data_files:
- split: train
path: MRI/annotations/reports_train.jsonl
- split: validation
path: MRI/annotations/reports_val.jsonl
- split: test
path: MRI/annotations/reports_test.jsonl
- config_name: mri_vqa
data_files:
- split: train
path: MRI/annotations/vqa_train.jsonl
- split: validation
path: MRI/annotations/vqa_val.jsonl
- split: test
path: MRI/annotations/vqa_test.jsonl
---
# ViMed-Rad sample release
This repository holds public samples of a de-identified Vietnamese radiology dataset of images, reports, and VQA annotations.
## MRI (brain): subset 100 / 20 / 20
`MRI/` contains 100 training, 20 validation, and 20 test brain MRI studies. They were drawn at random (seed 0) from the predefined splits, and their sample IDs match the full release. Each study keeps all of its diagnostic series as unpreprocessed NIfTI and is paired with a Vietnamese report and VQA annotations.
### Contents
- `MRI/images/<split>/<sample_id>/series_<NN>_<protocol>.nii.gz`: one NIfTI-1 volume per acquired or scanner-derived series, as converted from DICOM (no resampling, cropping, registration, skull-stripping, or intensity normalisation). Series are renumbered per study; `<protocol>` is the scanner protocol name (for example `t2_tse_dark-fluid_tra`, `t1_mprage_sag_p2_iso`, `ep2d_diff_..._ADC`).
- `MRI/annotations/reports_<split>.jsonl`: one row per study with
- `report`: findings section (Vietnamese), the text used as the generation target;
- `impression`: conclusion section (Vietnamese);
- `series`: file, protocol name, array shape, voxel spacing (mm), and `defaced`;
- `report_truncated_at_source`: true when the findings text was cut at 1,024 characters in the source export;
- `contrast_enhanced`: whether the exam was ordered with contrast (null if unknown);
- `model_input_sequences_available`: FLAIR / T1c / T1 / T2 availability used by the reference model.
- `MRI/annotations/vqa_<split>.jsonl`: `conversation` (human/gpt turns) and structured `vqa_items` (multiple-choice items keep `choices` and `answer_letter`).
- `MRI/dataset_card.json`: split counts, defacing summary, and preparation details.
| Split | Studies | Series files | VQA items |
|---|---:|---:|---:|
| train | 100 | 1,053 | 1,297 |
| validation | 20 | 191 | 165 |
| test | 20 | 211 | 234 |
| **total** | **140** | **1,455** | **1,696** |
### Privacy preparation
Source case identifiers, PACS URLs, source directory names, DICOM series numbers, and all DICOM metadata (dates, station, physicians, UIDs, demographics) were excluded. NIfTI header text fields and extensions were blanked, and gzip containers carry no original filename or timestamp. Scout/localizer series were removed. Report and VQA text passed automated redaction of URLs, e-mails, phone numbers, dates, and long numeric identifiers. The source-to-release ID mapping is not included.
3-D volumes with all voxel spacings at or below 1.5 mm were defaced. The pydeface template was registered to one 3-D T1 volume per study, and its face mask was transferred to the other 3-D series of the same session through scanner coordinates. Only voxels inside the face mask were set to zero; all other voxel values are unchanged. 2-D multi-slice series were not defaced. Each series entry records whether it was defaced.
## Responsible use
This release is intended for non-commercial research. Users must not attempt to re-identify individuals, reconstruct facial features, or link samples to external clinical data.
|