electricsheepafrica/Africa-Blood-Cell-Images-and-EHR-for-Cancer-Detection

Dataset Card: Blood cell images for cancer detection (with Synthetic Hematology EHR, Bundled Parquet)

Summary

This dataset, "Blood cell images for cancer detection," bundles hematology smear images and synthetic, Africa-aware hematology EHR per image into a single Parquet optimized for multimodal ML. Each image captures detailed cellular morphology under standardized conditions, focusing on both normal and abnormal blood cells.

Each row contains:

• image struct with {bytes, path}
• label (int) and label_name (string; inferred from folder)
• ehr_json (full EHR record as JSON string)
• Flattened EHR summary fields (demographics, triage, provider, inferred diagnosis)
• Complete blood count (CBC) and differential fields tailored to the image class

All EHR data are fully synthetic and non-identifiable.

This release contains 5,000 images (.jpg), with 1,000 images per class.

Technical Specifications (declared)

• Resolution: ≥ 1024×1024 pixels
• Staining: Wright–Giemsa
• Magnification: 100× oil immersion (≈1000× total)
• Color: 24-bit RGB
• Multiple focal planes per sample

Quality Measures (declared)

• Expert hematopathologist validation
• Standardized imaging conditions
• Multiple samples per cell type
• Detailed preparation documentation
• Complete technical metadata

Clinical Applications (intended)

• Normal vs. abnormal cell differentiation
• Leukemia subtype identification
• Disease progression monitoring
• Early detection screening
• Treatment response assessment

Image Annotations Include (scope)

• Nuclear patterns and contours
• Cytoplasmic features
• Nucleoli presence
• Cell measurements
• Abnormal inclusions (e.g., Auer rods)

Machine Learning Capabilities

• Automated cell classification
• Quantitative feature analysis
• Differential counting
• Morphological abnormality detection

2) Classes and Counts

Discovered from folder names under Blood Cell images for Cancer detection/:

• basophil: 1,000
• erythroblast: 1,000
• monocyte: 1,000
• myeloblast: 1,000
• seg_neutrophil: 1,000

Total rows: 5,000 (1,000 per class)

3) Schema (Bundled Parquet)

• image: struct<bytes: binary, path: string>
• label: int32
• label_name: string
• ehr_json: large_string
• EHR summaries (examples):
  • ehr_age, ehr_gender, ehr_country, ehr_urban_rural
  • ehr_facility_level, ehr_insurance_type, ehr_primary_language
  • ehr_time_to_presentation_days, ehr_triage_acuity, ehr_provider_specialty
  • ehr_symptoms, ehr_medications, ehr_diagnosis_suspected
• Labs (flattened):
  • lab_wbc, lab_hemoglobin, lab_platelets
  • lab_neutrophils_pct, lab_lymphocytes_pct, lab_monocytes_pct, lab_eosinophils_pct, lab_basophils_pct, lab_blasts_pct

The full EHR JSON schema is documented inline in the generator and aligns with typical CBC + differential and basic encounter context.

EHR Synthesis (Class-Conditional)

Hematology-aware logic infers plausible trends:

• myeloblast → elevated blasts_pct (AML suspected), thrombocytopenia, lower Hb
• basophil → basophilia, higher WBC/platelets (CML suspected)
• monocyte → monocytosis (CMML or infection suspected)
• erythroblast → anemia trend with nucleated RBC context (hemolysis/MDS suspected)
• seg_neutrophil → neutrophilia (bacterial infection suspected)

Demographics reflect Africa context (country, urban/rural, facility level, insurance, language) with longer presentation delays in rural/primary settings.

Coherence Audit (Passed)

Global delays (mean days):

• rural: 21.24
• urban: 7.85
• primary: 17.48
• tertiary: 9.70

Per-class means and fractions:

• basophil (n=1000):
  • hb=11.51, wbc=20.50, plt=379.1, neut=60.2, lymph=22.2, mono=6.7, eos=3.2, baso=6.5, blasts=1.1
  • frac_blasts>=20%=0.000, frac_baso>=2%=0.999, frac_neut>=65%=0.231, frac_mono>=10%=0.093
• erythroblast (n=1000):
  • hb=9.06, wbc=8.04, plt=241.2, neut=53.7, lymph=32.4, mono=8.6, eos=3.2, baso=1.1, blasts=1.1
  • frac_blasts>=20%=0.000, frac_baso>=2%=0.052, frac_neut>=65%=0.005, frac_mono>=10%=0.311
• monocyte (n=1000):
  • hb=11.02, wbc=12.24, plt=246.6, neut=51.2, lymph=22.6, mono=20.4, eos=3.1, baso=1.0, blasts=1.6
  • frac_blasts>=20%=0.000, frac_baso>=2%=0.052, frac_neut>=65%=0.021, frac_mono>=10%=0.996
• myeloblast (n=1000):
  • hb=9.54, wbc=18.45, plt=74.7, neut=23.0, lymph=20.7, mono=7.5, eos=2.1, baso=1.1, blasts=45.6
  • frac_blasts>=20%=1.000, frac_baso>=2%=0.076, frac_neut>=65%=0.000, frac_mono>=10%=0.213
• seg_neutrophil (n=1000):
  • hb=12.03, wbc=13.04, plt=282.1, neut=76.7, lymph=14.5, mono=4.4, eos=2.9, baso=1.0, blasts=0.5
  • frac_blasts>=20%=0.000, frac_baso>=2%=0.025, frac_neut>=65%=0.995, frac_mono>=10%=0.000

These align with intended clinical logic and thresholds (e.g., blasts in myeloblast, basophils in basophil, neutrophils in seg_neutrophil, anemia in erythroblast), and rural/primary delays > urban/tertiary.

Ethics and Limitations

• Fully synthetic EHR; no PHI.
• Trends are plausibility-driven and class-conditional; not diagnostic or epidemiologic evidence.
• Africa profile encodes general patterns (delays, access) and may not capture country-specific nuances.
• Rule-based generation cannot fully replicate real-world correlations.

License and Contact

• See repository LICENSE. If absent, assume “all rights reserved.”
• For questions or calibration requests, open an issue in this repository.