MS-CXR-T (Temporal Sentences and Image Pairs)¶

Modality: CXR | Format: JPG (MIMIC-CXR-JPG) | Dim: 2D | Labels: temporal progression (stable/improving/worsening) for 5 findings

Overview¶

Temporal benchmark built on top of MS-CXR. 1,045 longitudinal image pairs from MIMIC-CXR-JPG, where each row links a current CXR to a prior CXR from the same patient. Each pair is annotated with a progression label (stable / improving / worsening) for up to 5 pathology findings.

1,045 image pairs — one current image + one prior image per row
5 findings — consolidation, edema, pleural effusion, pneumonia, pneumothorax
Progression encoded as integer — -1 = improving, 0 = stable, 1 = worsening, NaN = not annotated
Label quality — each finding also has a quality column: one_expert / multiple_experts / disagreement / NaN

Source (PhysioNet — requires MIMIC credentialed access): - https://physionet.org/content/ms-cxr-t/

Images: MIMIC-CXR-JPG 2.0.0 — same image root as MS-CXR.

Download¶

# Download from PhysioNet
wget -r -N --no-parent -np \
  https://physionet.org/content/ms-cxr-t/1.0.0/ \
  -P ./ms-cxr-t/

# Images are from MIMIC-CXR-JPG 2.0.0 (same base_image_dir as MS-CXR)

Expected layout¶

<base_image_dir>/           ← MIMIC-CXR-JPG 2.0.0 root
  files/
    p10/
      p10002428/
        s55758034/
          3bea0373-....jpg
        ...

MS_CXR_T_temporal_image_classification_v1.0.0.csv   ← csv_path

Progression columns¶

Column	Values	Description
`consolidation_progression`	-1, 0, 1, NaN	Improving / stable / worsening
`edema_progression`	-1, 0, 1, NaN
`pleural_effusion_progression`	-1, 0, 1, NaN
`pneumonia_progression`	-1, 0, 1, NaN
`pneumothorax_progression`	-1, 0, 1, NaN

Quality columns¶

Column	Values	Description
`consolidation_label_quality`	`one_expert`, `multiple_experts`, `disagreement`, NaN
`edema_label_quality`	same
`pleural_effusion_label_quality`	same
`pneumonia_label_quality`	same
`pneumothorax_label_quality`	same

Extra metadata columns¶

Column	Type	Description
`previous_image_path`	`str`	Relative path to the prior-visit image
`previous_study_id`	`str`	Study ID of the prior visit

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes*	`None`	MIMIC-CXR-JPG 2.0.0 root (contains `files/`)
`csv_path`	`str`	Yes*	`None`	Path to `MS_CXR_T_temporal_image_classification_v1.0.0.csv`
`output_previous`	`bool`	No	`False`	Include prior-visit image under `"previous_img"`

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Include `"cls"` tensor in data dict
`output_mask`	`bool`	No	`False`	Include `"mask"` in data dict
`output_report`	`bool`	No	`False`	Include `"report"` in data dict
`output_bbox`	`bool`	No	`False`	Include `"bbox"` and `"bbox_labels"` in data dict
`transform`	Compose	No	standard 2-D 224 px	MONAI Compose transform
`cache_dir`	`str`	No	`"./cache"`	MONAI cache directory. `None` disables
`dtype`	`torch.dtype`	No	`torch.bfloat16`	Output tensor dtype
`harmonized_df`	`pd.DataFrame`	No	`None`	Pre-built harmonized DataFrame
`harmonizer`	harmonizer	No	`None`	Pre-instantiated harmonizer
`harmonizer_path`	`str`	No	`None`	Path to saved harmonizer pickle

* Required unless harmonizer_path or harmonized_df is provided.

Dataset constructor¶

from radharmony.dataset import MSCXRTDataset

ds = MSCXRTDataset(
    base_image_dir="/data/mimic-cxr-jpg/2.0.0/",
    csv_path="/data/ms-cxr-t/MS_CXR_T_temporal_image_classification_v1.0.0.csv",
    cache_dir="./cache",
)

sample = ds.get_datasets()[0]
print(sample["img"].shape)   # torch.Size([3, 224, 224]) — current visit

Access progression labels and prior images via the harmonized DataFrame:

df = ds.get_harmonized_df()
print(df[["study_id", "previous_study_id",
          "pleural_effusion_progression",
          "pleural_effusion_label_quality"]].head(4))
#    study_id  previous_study_id  pleural_effusion_progression  pleural_effusion_label_quality
# 0  55758034          50414267                          0.0                    multiple_experts
# 1  57375967          54276838                          1.0                    one_expert
# ...

Harmonizer¶

from radharmony.harmonizer import MSCXRTHarmonizer

h = MSCXRTHarmonizer(
    csv_path="/data/ms-cxr-t/MS_CXR_T_temporal_image_classification_v1.0.0.csv",
)
df = h.harmonize()
print(df.shape)           # (1045, 15)
print(df.columns.tolist())
# ['patient_id', 'study_id', 'image_path', 'previous_image_path', 'previous_study_id',
#  'consolidation_progression', 'edema_progression', 'pleural_effusion_progression',
#  'pneumonia_progression', 'pneumothorax_progression',
#  'consolidation_label_quality', 'edema_label_quality', 'pleural_effusion_label_quality',
#  'pneumonia_label_quality', 'pneumothorax_label_quality']

Harmonizer notes¶

No LABEL_COLS — this dataset has no binary classification targets; progression labels live in _progression columns.
Progression encoding — raw strings ("improving" / "stable" / "worsening") are mapped to integers (-1 / 0 / 1). Missing annotations remain NaN.
image_path — derived from dicom_id in the CSV via "files/" + dicom_id + ".jpg". Format: files/p10/pXXXXXXXX/sYYYYYYYY/<hash>.jpg.
previous_image_path — same derivation applied to previous_dicom_id.
Images are shared with MIMIC-CXR-JPG — both current and previous images live under the same base_image_dir.