Phase Grounding: MIMIC-CXR Label Extraction

Extracting TAIX-style ordinal labels and laterality from MIMIC-CXR free-text reports using local regex NLP — no external API calls, no data leaves the machine.

Goal

TAIX-Ray uses a 5-point ordinal severity scale (0–4) with laterality (left/right/bilateral) for each finding. MIMIC-CXR-JPG ships CheXpert binary labels that cannot be converted to TAIX ordinals. Instead, the free-text radiology reports contain the severity and laterality information — we extract it directly.

Output schema

Matches the TAIX harmonized label columns:

Column	Type	Description
subject_id	int	MIMIC patient identifier
study_id	int	MIMIC study identifier
atelectasis	float	0–4 severity (collapsed), NaN if absent from report
atelectasis_left	float	Left-side severity
atelectasis_right	float	Right-side severity
pleural_effusion	float	0–4 severity (collapsed)
pleural_effusion_left	float
pleural_effusion_right	float
pulmonary_opacities	float	0–4 severity (collapsed)
pulmonary_opacities_left	float
pulmonary_opacities_right	float
heart_size	float	0–4 cardiomegaly severity
pulmonary_congestion	float	0–4 vascular congestion severity

Severity scale:

Score	Meaning
0	Absent / resolved / normal
1	Trace / minimal / tiny
2	Mild / slight
3	Moderate
4	Large / severe / massive
NaN	Finding not mentioned in the report

Extraction pipeline

Script: /path/to/phase_grounding/extract_mimic_labels.py

Step 1 — section extraction

Reports are split into FINDINGS and IMPRESSION sections using a regex that handles the common section header patterns. FINDINGS is preferred; IMPRESSION is the fallback; the full report text is the last resort.

Step 2 — finding detection

Each finding has a compiled regex pattern:

Finding	Patterns
atelectasis	atelectasis, atelectatic, discoid atelectasis, subsegmental atelectasis, linear atelectasis, basilar atelectasis
pleural_effusion	pleural effusion, pleural fluid, hydrothorax, effusion
pulmonary_opacities	opacity, infiltrate, consolidation, airspace disease, alveolar/focal/patchy opacit
heart_size	cardiomegaly, enlarged heart, cardiac silhouette, heart size, heart is … enlarg*
pulmonary_congestion	pulmonary edema, vascular congestion/engorgement, interstitial edema, prominent vascular markings

Step 3 — severity scoring

Within each sentence containing a finding, the severity patterns fire (in order) and the highest-severity match wins. If a sentence mentions a finding but has no severity qualifier, a default of 2 (mild) is assigned.

Negation ("no ", "without ", "absent", "free of") and normal-range modifiers ("normal", "unremarkable", "within normal limits", "not enlarged") both score 0.

Step 4 — laterality

Each positive sentence is examined for: - "bilateral", "bibasilar", "basal", "both lung" → both sides - "left", "left-sided" → left only - "right", "right-sided" → right only

The collapsed score is max(left, right). Findings without a side word (e.g. "cardiomegaly") get no laterality columns.

Running the extraction

cd /path/to/phase_grounding
uv run --with pyarrow --with pandas --with tqdm python3 extract_mimic_labels.py

Processes 227,835 MIMIC-CXR studies in ~5 minutes on a single CPU core. Output:

data/mimic_taix_labels.parquet

Descriptive statistics (227,835 studies)

Finding coverage (non-null count = finding mentioned in report)

Finding	Studies with label	% coverage
atelectasis	~145,000	64%
pleural_effusion	~115,000	50%
pulmonary_opacities	~125,000	55%
heart_size	~170,000	75%
pulmonary_congestion	~70,000	31%

Severity distribution (atelectasis example)

Score	Count	% of labelled
0 (absent)	~42,000	29%
1 (trace)	~8,000	6%
2 (mild)	~72,000	50%
3 (moderate)	~15,000	10%
4 (large)	~8,000	6%

Laterality (atelectasis)

Radiologists predominantly use "bibasilar" rather than left/right — most labels are bilateral or NaN; explicit left-only / right-only are a minority.

Using extracted labels with RadHarmony

The parquet file contains subject_id and study_id matching MIMIC-CXR-JPG. Merge with the harmonized MIMIC-CXR-JPG DataFrame to add TAIX-style labels:

import pandas as pd
from radharmony.harmonizer.mimic_cxr.mimic_cxr_jpg import MIMICCXRJPGHarmonizer
from radharmony.dataset import MIMICCXRJPGDataset

MIMIC_BASE = "/path/to/MIMIC_CXR/physionet.org/files/mimic-cxr-jpg/2.0.0/"
LABELS_PARQUET = "/path/to/phase_grounding/data/mimic_taix_labels.parquet"

# 1. Harmonize MIMIC-CXR-JPG (builds image_path, patient_id, study_id, CheXpert labels …)
h = MIMICCXRJPGHarmonizer(
    csv_path=f"{MIMIC_BASE}mimic-cxr-2.0.0-metadata.csv.gz",
    label_csv_path=f"{MIMIC_BASE}mimic-cxr-2.0.0-chexpert.csv.gz",
)
mimic_df = h.harmonize()
# mimic_df["patient_id"] is subject_id as string, "study_id" as string

# 2. Load extracted severity labels
pg_df = pd.read_parquet(LABELS_PARQUET)
pg_df["patient_id"] = pg_df["subject_id"].astype(str)
pg_df["study_id"]   = pg_df["study_id"].astype(str)

# 3. Merge — left join keeps all MIMIC rows, NaN where no report label exists
merged = mimic_df.merge(
    pg_df.drop(columns="subject_id"),
    on=["patient_id", "study_id"],
    how="left",
)

# 4. Load dataset with the merged DataFrame (no re-harmonization)
ds = MIMICCXRJPGDataset(
    base_image_dir=f"{MIMIC_BASE}files/",
    output_cls=True,
    harmonized_df=merged,
)

sample = ds.get_datasets()[0]
print(sample["img"].shape)   # torch.Size([1, 224, 224])
# Access severity labels directly from the harmonized DataFrame:
print(merged[["study_id", "atelectasis", "atelectasis_left", "atelectasis_right",
              "pleural_effusion", "heart_size"]].head())

Eval app

A standalone Gradio app (port 7861) lets you test the regex extraction on any report text:

cd /path/to/phase_grounding
uv run --with gradio --with pandas --with pyarrow python3 eval_app.py

Open http://localhost:7861: - Live Demo tab — paste any report, see extracted labels with severity scores and laterality - Dataset Browser tab — browse the 227,835 extracted labels, filter by finding / side / severity range

Known limitations

• Severity default of 2 (mild) — when a finding is mentioned without a clear severity qualifier (e.g. "atelectasis is present"), we assign 2. This over-represents mild severity.
• Bibasilar dominance — radiologists often write "bibasilar atelectasis" without specifying left vs right, making the _left/_right columns NaN even when severity > 0.
• Negation scope — negation is sentence-level only. "No pleural effusion. Small right pleural effusion." correctly extracts score 1 for the right side. But complex negation like "Effusion previously present is now resolved" may not be captured accurately.
• Normal heart in context — "cardiac silhouette is within normal limits" → score 0 (correct). But "stable cardiomegaly" — "stable" fires a score-0 pattern; however, the finding regex also fires with no severity, yielding default 2. Both patterns coexist — the final score is max(0, 2) = 2 (mild cardiomegaly), which is reasonable.