IDC TCGA-ESCA: Esophageal Carcinoma whole-slide microscopy, CT (185 patients): sample notebook

A worked notebook for IDC TCGA-ESCA: Esophageal Carcinoma whole-slide microscopy, CT (185 patients), with 7 code cells and the output each one printed. On Scigantic it opens in a hosted JupyterLab with the dataset already mounted, so nothing is downloaded first.

Provider
NCI Imaging Data Commons; TCGA-ESCA contributors; tables and sample by Scigantic
License
Per series (series.parquet license_short_name): CC BY 3.0 522, CC BY 4.0 312. Cite the source DOI(s) and the IDC.
Size
8 parquet tables plus 29 sample DICOM files (24 MB)
Stored at
Amazon S3, s3://scigantic-idc-open/tcga_esca (us-east-1)
Access
Free and public. No account with the data provider is needed.
DOI
10.5281/zenodo.12690012

The notebook

TCGA-ESCA on the Imaging Data Commons: 185 patients, 834 series, 247.38 GB

Esophageal Carcinoma imaging of the esophagus from the NCI Imaging Data Commons (IDC). Series by modality: SM 396, SEG 312, CT 126.

This mount is the collection's index plus a real sample, not the collection: series.parquet (one row per DICOM series, with the public S3 location of each), studies.parquet, patients.parquet, the IDC clinical tables, and sample/ folders holding one CT series (CC BY 3.0); one SEG series (CC BY 4.0); one SM slide (thumbnail + two smallest pyramid levels) (CC BY 3.0). Anything else is one anonymous S3 read away. README.md and BUILD_REPORT.json at the mount root say exactly how this was built (idc-index, IDC v25).

The code below reads all of it with the scigantic-nci package (pip install "scigantic-nci[dicom]", already installed in this image): from scigantic_nci import idc, then calls such as idc.series("tcga_esca"), idc.read_sample and idc.pull_series. Run anywhere else, the same calls read s3://scigantic-idc-open/tcga_esca/ anonymously.

import os, json
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import pydicom
from scigantic_nci import idc, NciError
plt.rcParams['figure.dpi'] = 72  # keeps the saved figures small enough for the notebook preview
try:
    from matplotlib_inline.backend_inline import set_matplotlib_formats
    set_matplotlib_formats('jpeg')  # slide renders are photographic; JPEG is a fifth the size of PNG here
except ImportError:
    pass

C = 'tcga_esca'
# scigantic_nci reads the mounted archive when its BUILD_REPORT.json names this collection, else s3://scigantic-idc-open/<collection>/ anonymously.
M = os.environ.get('SCIGANTIC_MOUNT_PATH', '/mnt/archive')
on_mount = os.path.exists(f'{M}/BUILD_REPORT.json') and json.load(open(f'{M}/BUILD_REPORT.json')).get('collection_id') == C
print('reading', C, 'from', M if on_mount else f's3://scigantic-idc-open/{C}/')

report = idc.report(C)
series = idc.series(C)  # one row per DICOM series, with its public S3 location
patients = idc.patients(C)
studies = idc.studies(C)
print(report['collection_name'], '| IDC', report['idc_data_version'], '| built', report['built_at'][:10])
print(f"{len(patients):,} patients, {len(studies):,} studies, {len(series):,} series, {series.series_size_MB.sum()/1024:,.1f} GB in the IDC buckets\n")
by_mod = series.groupby('Modality').agg(series=('SeriesInstanceUID', 'count'), patients=('PatientID', 'nunique'), GB=('series_size_MB', lambda s: round(s.sum() / 1024, 2))).sort_values('series', ascending=False)
print(by_mod.to_string())
fig, ax = plt.subplots(figsize=(7, 3))
by_mod['patients'].plot.bar(ax=ax, color='#4a6fa5'); ax.set_ylabel('patients'); ax.set_title(f"{report['collection_name']}: patients per modality"); plt.tight_layout()

Output:

reading tcga_esca from /mnt/archive
TCGA-ESCA | IDC v25 | built 2026-11-17
185 patients, 202 studies, 834 series, 247.4 GB in the IDC buckets

          series  patients      GB
Modality                          
SM           396       185  236.66
SEG          312       156    0.03
CT           126        16   10.69

1. Studies and patients

A study is one imaging visit (all series acquired together); a patient may have several. studies.parquet collapses the series table to one row per study with the modalities present, patients.parquet to one row per patient.

print('Studies per patient:'); print(patients.n_studies.describe()[['min', '50%', 'max']].to_string())
print('\nModality combinations per study (top 8):'); print(studies.modalities.value_counts().head(8).to_string())
sex = patients.PatientSex.replace('', np.nan).dropna()
if len(sex): print('\nPatientSex:', sex.value_counts().to_dict())
ages = pd.to_numeric(series.PatientAge.dropna().astype(str).str.extract(r'(\d+)')[0], errors='coerce')
if ages.notna().sum() > 10: print(f'PatientAge at imaging: median {ages.median():.0f} y (IQR {ages.quantile(.25):.0f}-{ages.quantile(.75):.0f}) on {ages.notna().sum():,} series')
if 'BodyPartExamined' in series: print('\nBodyPartExamined:', series.BodyPartExamined.dropna().value_counts().head(6).to_dict())
print('\nLicenses (per series):', series.license_short_name.value_counts().to_dict())

Output:

Studies per patient:
min    1.0
50%    1.0
max    3.0

Modality combinations per study (top 8):
modalities
SEG, SM    156
SM          29
CT          17

PatientSex: {'M': 13, 'F': 3}
PatientAge at imaging: median 56 y (IQR 51-64) on 126 series

BodyPartExamined: {'ESOPHAGUS': 126}
...

2. Open the sample CT series

sample/CT_1.3.6.1.4.1.14519.5.2.1.6354.4026.981064781651495417235238573859/ is the complete series of patient TCGA-VR-A8EX (25 instances, 15.1 MB, CC BY 3.0), copied byte-for-byte from s3://idc-open-data/0c8683fb-3383-402f-a1cf-dce866388c47/*. One .dcm per slice; idc.read_sample sorts them on ImagePositionPatient[2] (InstanceNumber as a fallback) and stacks them, applying RescaleSlope/RescaleIntercept so values are Hounsfield units. The pydicom datasets are in vol.datasets.

try:
    vol = idc.read_sample(C, 'CT')  # Volume: instances sorted by ImagePositionPatient z then InstanceNumber, CT in Hounsfield units
except NciError as e:  # compressed transfer syntax with no decoder plugin in this image
    print(e, 'Skipping the render.')
    vol = None
if vol is not None:
    n = len(vol.datasets)
    if vol.array is not None and n >= 3:
        lo, hi = -160, 240  # soft-tissue window in Hounsfield units
        picks = [(vol.array[i], f'slice {i + 1}/{n}') for i in (n // 4, n // 2, 3 * n // 4)]
    else:  # 2D modalities or a series that mixes image sizes (scouts, crops): show instances one by one
        imgs = vol.images if vol.ragged else list(vol.array)
        picks = [(img, f'{d.get("SeriesDescription", "")} {img.shape[1]}x{img.shape[0]}') for img, d in zip(imgs[:3], vol.datasets[:3])]
        flat = np.concatenate([p[0].ravel() for p in picks]); lo, hi = np.percentile(flat, 1), np.percentile(flat, 99)
    shapes = sorted({im.shape for im in vol.images}) if vol.ragged else [vol.array.shape[1:]]
    d0 = vol.datasets[0]
    print(f'{vol.modality} {n} instances, image shapes {shapes} | spacing {d0.get("PixelSpacing") or d0.get("ImagerPixelSpacing")} mm, slice thickness {d0.get("SliceThickness")} mm | {d0.get("Manufacturer")} {d0.get("ManufacturerModelName")} | {d0.get("SeriesDescription")}')
    fig, axes = plt.subplots(1, len(picks), figsize=(3.6 * len(picks), 3.8))
    for ax, (img, label) in zip(np.atleast_1d(axes), picks):
        ax.imshow(img if img.ndim == 2 else np.clip((img - lo) / max(hi - lo, 1e-6), 0, 1), cmap='gray', vmin=lo, vmax=hi); ax.set_title(label, fontsize=9); ax.axis('off')  # RGB (ultrasound) is scaled to 0-1
    plt.suptitle(f"TCGA-ESCA: sample CT series, patient TCGA-VR-A8EX"); plt.tight_layout()

Output:

CT 25 instances, image shapes [(512, 512), (512, 592)] | spacing [0.939453125, 0.939453125] mm, slice thickness 4 mm | SIEMENS Emotion 6 (2007) | CORONAL PULMAO

3. The sample segmentation

sample/SEG_1.2.826.0.1.3680043.10.511.3.50778481665835016196838040503212988/ is a DICOM SEG series of patient TCGA-V5-A7RE (0.0 MB, CC BY 4.0): a stack of binary frames, one per segment and source slice, with SegmentSequence naming each segment and ReferencedSeriesSequence pointing at the image series it was drawn on. A few collections label ordinary image instances (one label map per slice) as SEG; the cell handles both. idc.sample_files lists the files; on the mount they are read in place, and elsewhere the same series is pulled from IDC with idc.pull_series. The source image series can be pulled from IDC the same way as any other row of series.parquet.

# scigantic_nci has no SEG reader, and read_sample decodes one frame per instance, which pydicom cannot do for some 1-bit SEG objects,
# so the SEG files are read whole with pydicom: in place on the mount (sample_files gives paths relative to the archive root), else pulled from IDC.
seg_sample = idc.samples(C).query("modality == 'SEG'").iloc[0]
if on_mount:
    seg_files = [os.path.join(M, f) for f in idc.sample_files(C, 'SEG')]
else:
    seg_files = idc.pull_series(series[series.SeriesInstanceUID == seg_sample.SeriesInstanceUID].iloc[0], '/tmp/idc_pull/' + seg_sample.SeriesInstanceUID)
seg_ds = [pydicom.dcmread(f) for f in seg_files]
seg = seg_ds[0]
if 'SegmentSequence' in seg:  # a DICOM Segmentation object
    print(seg.Modality, seg.get('SeriesDescription'), '|', int(seg.get('NumberOfFrames', 1)), 'frames of', seg.Rows, 'x', seg.Columns, '|', seg.get('ContentCreatorName'))
    print('segments:', [(int(x.SegmentNumber), str(x.SegmentLabel), str(x.get('SegmentAlgorithmType'))) for x in seg.SegmentSequence][:8])
    mask = seg.pixel_array
else:  # labelled SEG but stored as ordinary images: one label map per slice
    seg_ds.sort(key=lambda d: (float(d.ImagePositionPatient[2]) if 'ImagePositionPatient' in d else 0.0, int(d.get('InstanceNumber', 0))))
    print(seg.Modality, seg.get('SeriesDescription'), '|', len(seg_ds), seg.SOPClassUID.name, 'instances of', seg.Rows, 'x', seg.Columns, '(label maps, not a Segmentation object)')
    mask = np.stack([d.pixel_array for d in seg_ds])
    values, counts = np.unique(mask, return_counts=True)
    mask = mask != values[counts.argmax()]  # label maps carry a nonzero background value; everything else is labelled
ref = seg.ReferencedSeriesSequence[0].SeriesInstanceUID if 'ReferencedSeriesSequence' in seg else None
src = series[series.SeriesInstanceUID == ref]
if len(src):
    r = src.iloc[0]
    print(f'drawn on the {r.Modality} series of patient {r.PatientID}: {int(r.instanceCount)} instances, {r.series_size_MB:.1f} MB, {r.license_short_name}, {r.series_aws_url}')
mask = (mask[None] if mask.ndim == 2 else mask) > 0
i = int(mask.reshape(mask.shape[0], -1).sum(1).argmax())  # the frame with the most labelled pixels
print(f'{int(mask.sum()):,} labelled pixels over {mask.shape[0]} frames; densest frame {i + 1}')
fig, ax = plt.subplots(figsize=(3.8, 3.8)); ax.imshow(mask[i], cmap='gray'); ax.axis('off'); ax.set_title(f"TCGA-ESCA: sample SEG, frame {i + 1}", fontsize=9); plt.tight_layout()

Output:

SEG Stony Brook Inception-V4 Binary TIL Map | 1 frames of 143 x 154 | 
segments: [(1, 'TILs Present', 'AUTOMATIC')]
drawn on the SM series of patient TCGA-V5-A7RE: 4 instances, 60.3 MB, CC BY 3.0, s3://idc-open-data/84b32551-dfea-4579-b06a-9ca24855f337/*
132 labelled pixels over 1 frames; densest frame 1

4. Open the sample whole-slide image

Slide microscopy in IDC is DICOM WSI: every resolution level of the pyramid is its own instance, tiled into frames. sample/SM_1.3.6.1.4.1.5962.99.1.1917100131.582423934.1638299672675.2.0/ holds the thumbnail and the two smallest levels of one slide of patient TCGA-V5-AASX (the full slide is 9 MB in 3 instances). idc.read_sample lists the levels, lowest resolution first, and level(i) assembles one from its tiles (idc.assemble_level(path) does the same for any WSI instance on disk); idc.slides is the slide table with the stain flattened to text. wsidicom (installed here) reads arbitrary regions of a full slide.

slide = idc.read_sample(C, 'SM')  # SlideLevels: one entry per pyramid instance, lowest resolution first
for lv in slide.levels: print(f'{lv.width:6d} px wide  {lv.image_type}')
top = slide.levels[-1]
img = slide.level(top.index)  # tiles assembled row-major into one array
print(f'assembled {img.shape[1]} x {img.shape[0]} px from {top.n_frames} tiles of {top.rows} x {top.cols}' + (f', {top.um_per_px:.2f} um/px at this level' if top.um_per_px else ''))
stains = idc.slides(C)['stain'].dropna()  # sm_series joined to series, list-valued stain codes flattened to text
print('stains in the collection:', stains.value_counts().head(4).to_dict())
step = max(1, int(np.ceil(max(img.shape[:2]) / 700)))  # show at most ~700 px on the long side
shown = img[::step, ::step]
fig, ax = plt.subplots(figsize=(6, 6 * shown.shape[0] / shown.shape[1])); ax.imshow(shown, cmap=None if shown.ndim == 3 else 'gray'); ax.axis('off'); ax.set_title(f"TCGA-ESCA: sample slide, patient TCGA-V5-AASX"); plt.tight_layout()

Output:

   752 px wide  THUMBNAIL
  1991 px wide  VOLUME
  7967 px wide  VOLUME
assembled 7967 x 8135 px from 1156 tiles of 240 x 240, 0.50 um/px at this level
stains in the collection: {'hematoxylin stain, water soluble eosin stain': 396}

5. Clinical tables

IDC ships 1 clinical table(s) for this collection (tcga_esca_clinical; idc.clinical_tables lists them), keyed by dicom_patient_id = PatientID. clinical/dictionary.parquet (idc.clinical_dictionary) carries the human-readable label and the coded values for every column; coded columns store the code.

clin = idc.clinical(C, 'tcga_esca_clinical')
dictionary = idc.clinical_dictionary(C)
print(clin.shape, 'rows x cols in', 'tcga_esca_clinical')
print(dictionary[dictionary.short_table_name == 'tcga_esca_clinical'][['column', 'column_label']].head(15).to_string(index=False))
joined = patients.merge(clin, left_on='PatientID', right_on='dicom_patient_id', how='left')
print(f'\n{joined.dicom_patient_id.notna().sum()} of {len(patients)} imaged patients have a clinical row')

Output:

(185, 74) rows x cols in tcga_esca_clinical
                       column                    column_label
             dicom_patient_id idc_provenance_dicom_patient_id
                 source_batch     idc_provenance_source_batch
                  case_gdc_id                     case_gdc_id
                 case_barcode                    case_barcode
                 program_name                    program_name
                 disease_code                    disease_code
           project_short_name              project_short_name
             age_at_diagnosis                age_at_diagnosis
   age_began_smoking_in_years      age_began_smoking_in_years
anatomic_neoplasm_subdivision   anatomic_neoplasm_subdivision
                 batch_number                    batch_number
                          bcr                             bcr
                          bmi                             bmi
...

6. Pull any series from IDC on demand

The collection is 247.38 GB in s3://idc-open-data (and idc-open-data-cr for CC BY-NC series); it is not on this mount. Every row of series.parquet has aws_bucket and crdc_series_uuid; the objects under that prefix are the series' instances. idc.pull_series(row, dest) takes such a row and downloads the instances anonymously, so this needs no credentials. Check series_size_MB before pulling: a pod has a few GB of local disk.

cand = idc.series(C, modality='SM', max_size_mb=20).sort_values('series_size_MB')  # filters pushed down to parquet
row = cand.iloc[len(cand) // 2] if len(cand) else series.sort_values('series_size_MB').iloc[0]
print(row.PatientID, row.Modality, int(row.instanceCount), 'instances,', round(idc.series_size(row), 1), 'MB ->', row.series_aws_url)
print('IDC viewer:', idc.viewer_url(row))
files = idc.pull_series(row, '/tmp/idc_pull')  # anonymous read of s3://<aws_bucket>/<crdc_series_uuid>/, returns the local paths
d = pydicom.dcmread(files[0], stop_before_pixels=True)
print(len(files), 'files downloaded;', d.Modality, d.get('SeriesDescription'), d.get('Rows'), 'x', d.get('Columns'))

Output:

TCGA-V5-A7RE SM 3 instances, 18.5 MB -> s3://idc-open-data/5db693b0-2064-45bc-a609-b87ca31c6f10/*
IDC viewer: https://viewer.imaging.datacommons.cancer.gov/viewer/2.25.37553261986983463509652718466687957324?seriesInstanceUID=1.3.6.1.4.1.5962.99.1.1933429942.1958224744.1638316002486.2.0
3 files downloaded; SM Frozen HE NT TS1 256 x 256

Try this in the AI assistant

The chat panel is prefilled with the first of these.

Licenses are per series (license_short_name in series.parquet). Cite the collection DOI(s) in README.md and the IDC papers (Fedorov et al. 2021, 2023) when you publish.

Run it

Open this notebook to run it, see the dataset page for the full description, or browse the full catalog.