Skip to content

Datasets

RadHarmony supports 44 dataset configurations across 35 underlying datasets. Click a dataset name for the full page (constructor args, harmonizer notes, label columns).

Inventory

Dataset Modality Dim Labels Outputs available Page
CheXpert CXR 2D 14 pathologies cls →
CheXpert-Plus CXR 2D 14 pathologies cls, report →
CheXlocalize CXR 2D 14 pathologies cls, mask →
VQA-RAD CXR/CT/MRI 2D n/a (VQA) question, answer →
MIMIC-CXR (DICOM) CXR 2D 14 pathologies cls, report →
MIMIC-CXR-JPG CXR 2D 14 pathologies cls →
MIMIC-CXR-JPG (Test) CXR 2D 14 pathologies cls →
ChestX-ray14 CXR 2D 15 pathologies cls →
ChestX-ray14 (BBox) CXR 2D 15 pathologies cls, bbox →
PadChest CXR 2D 193 findings cls, report (Spanish) →
ReXGradient-160K (Train) CXR 2D none report →
ReXGradient-160K (Valid) CXR 2D none report →
ReXGradient-160K (Test) CXR 2D none report →
VinDr-CXR (Train) CXR 2D 28 findings cls, bbox →
VinDr-CXR (Test) CXR 2D 28 findings cls, bbox →
VinDr-PCXR CXR 2D 15 conditions (pediatric) cls, bbox →
SIIM-ACR Pneumothorax CXR 2D pneumothorax cls, mask →
SIIM COVID-19 CXR 2D 4 appearance classes cls, bbox →
RSNA Pneumonia CXR 2D 3 classes cls, bbox →
RSNA Pneumonia (Kaggle) CXR 2D 3 classes cls, bbox →
RSNA PE Detection beta CT 3D 13 PE labels cls →
RSNA Pediatric Bone Age beta Radiograph 2D age (regression) reg →
RSNA 2024 Lumbar Spine beta MRI 3D 75 severity cols cls, bbox →
CT-RATE beta CT 3D 18 findings cls →
RAD-ChestCT beta CT 3D 84 findings cls, mask, report, bbox →
Montgomery County CXR CXR 2D TB classification cls, mask →
Shenzhen Hospital CXR CXR 2D tuberculosis (binary) cls, mask, report →
OpenI IU CXR CXR 2D 8 findings (CheXpert-compat) cls, report →
OpenI IU CXR (DICOM) CXR 2D 8 findings (CheXpert-compat) cls, report →
TAIX-Ray (512px) CXR 2D 8 findings (binary/ordinal) cls →
TAIX-Ray (original) CXR 2D 8 findings (binary/ordinal) cls →
RSNA 2022 Cervical Spine beta CT 3D 8 fracture labels cls, mask →
RSNA 2022 Cervical Spine (BBox) beta CT 3D 8 fracture labels cls, bbox →
RSNA 2023 Abdominal Trauma beta CT 3D 14 injury labels cls →
BRAX (DICOM) CXR 2D 14 pathologies cls →
BRAX (PNG) CXR 2D 14 pathologies cls →
RANZCR CLiP CXR 2D 11 catheter/line labels cls, mask →
GEMeX-VQA CXR 2D n/a (VQA over MIMIC-CXR-JPG) question, answer, bbox →
MIMIC-Ext-CXR-QBA CXR 2D n/a (VQA over MIMIC-CXR) question, answer →
ROCO Multimodal 2D n/a (captioning) question (empty), answer, keywords →
EmoryCXR v2 CXR 2D 14 pathologies cls, report →
Emory CHORUS (X-ray subset) CXR 2D none (image-only) img →
MS-CXR CXR 2D 8 findings + phrase grounding cls, bbox →
MS-CXR-T CXR 2D temporal progression (5 findings) previous_img →

Finding a dataset's label columns

Every dataset class exposes LABEL_COLS and REG_COLS as class attributes:

from radharmony.dataset import CheXpertDataset, RSNABoneAgeDataset

print(CheXpertDataset.LABEL_COLS)   # ['atelectasis', 'cardiomegaly', ...]
print(RSNABoneAgeDataset.REG_COLS)  # ['age_months']

The cls tensor in a data dict has one value per entry in LABEL_COLS, in sorted (alphabetical) order — not necessarily the order they appear in the class definition.

Output flags

Each dataset supports a subset of output flags. Passing a flag that the dataset does not support raises a warning and the key is absent from the data dict.

Flag Data dict key Description
output_cls=True "cls" Multi-label classification tensor
output_mask=True "mask" Segmentation mask tensor
output_bbox=True "bbox", "bbox_labels" List of [dim0_min, dim0_max, dim1_min, dim1_max] fractional boxes
output_report=True "report" Free-text radiology report string
output_reg=True "reg" Regression target tensor

base_image_dir convention

base_image_dir is the deepest common ancestor directory shared by all images in the dataset — not the competition root. image_path values are relative to it and may include subdirectory components (e.g. patient_id/study_id/image.dcm for datasets with patient/study folder structure). For each dataset, the wiki page above documents the exact expected layout and an example path.

For datasets with separate train/test splits, each split takes its own image directory:

Dataset Train base Test base
CheXpert train/ valid/
ChestX-ray14 CXR14 root (images_001/ … images_012/ siblings) same
VinDr-CXR train/ test/
SIIM-ACR-PTX dicom-images-train/ dicom-images-test/
SIIM COVID-19 train/ test/
RSNA Pneumonia (Kaggle) stage_2_train_images/ stage_2_test_images/
RSNA PE Detection train/ test/
RSNA Cervical Spine train_images/ test_images/
RSNA Abdominal Trauma train_images/ test_images/
RSNA Lumbar Spine train_images/ test_images/
RSNA Bone Age boneage-training-dataset/ Bone Age Validation Set/
RANZCR CLiP train/ test/
ReXGradient-160K deid_png/ (train split) deid_png/ (test split — JSON differs per split, image tree shared)

Structural exceptions — some datasets have images split across multiple sibling subdirs inside base_image_dir, where image_path carries the subdir name. Examples:

  • MIMIC-CXR (DICOM) and MIMIC-CXR-JPG: base_image_dir is the files/ directory; images are grouped under p10/, p11/, … patient-group subdirectories, so image_path takes the form p{group}/p{patient_id}/s{study_id}/{image}.dcm (or .jpg).
  • ChestX-ray14: base_image_dir is the CXR14 root; images are split across its images_001/ … images_012/ sibling children.
  • PadChest: base_image_dir is the images/ directory; images are split across its 0/, 1/, …, 50/, 54/ sibling children (slots 51/52/53 don't exist in the official distribution).
  • RSNA Bone Age (val): base_image_dir is Bone Age Validation Set/; images are split across its boneage-validation-dataset-1/ and -2/ sibling children. (Train has no exception — its base_image_dir is the flat image directory.)

When adding a new dataset, follow the base_image_dir convention: no hardcoded subdirectory prefix in image_path that could be folded into base_image_dir instead.