Perle-ai/multimodal-ct-radiology-reports

Perle AI Multi-phase CECT and CT with Radiology Reports

Summary

A de-identified CT dataset from Perle AI, paired with the original radiology reports. It supports work on multi-modal medical imaging: phase or pathology classification, report generation from images, and visual question answering.

The release has three configurations:

10 subjects, 171 series, 32,225 DICOM instances total.

Supported tasks

• —Image classification: phase or pathology label per study.
• —Image-to-text: report generation from a CT volume.
• —Visual question answering: answer questions about a CT given the paired report.
• —Segmentation: no masks ship today; consumers can supply their own.

Languages

Reports are in clinical English. Some abbreviations follow South-Asian hospital conventions. No translation has been applied.

Dataset structure

multimodal-ct-radiology-reports/
├── README.md
├── LICENSE
├── DATA_USE_AGREEMENT.md
├── CHANGELOG.md
├── cect_3phase/
│   ├── manifest.csv
│   ├── metadata_catalogue.csv
│   └── PA*/PA*.txt + PA*/ST*/SE*/IM*.dcm
├── ct_general/
│   ├── manifest.csv
│   └── PA*_ct_*/<report>.txt + PA*_ct_*/ST*/SE*/IM*
└── gb_carcinoma/
    ├── manifest.csv
    ├── cancer_report.txt
    └── PA0/ST*/SE*/IM*

manifest.csv schema

Data and acquisition

Curated by Perle AI from clinical CT studies acquired in 2026. All studies were captured on a Siemens SOMATOM go.Now 32-slice MDCT with 0.7–0.8 mm axial acquisition reconstructed to 5 mm axial / coronal / sagittal MPR. The cect_3phase cohort follows a standard non-contrast / arterial / portal-venous protocol.

Inclusion: clinical CT studies with a completed English-language radiology report. Exclusion: studies without a report, studies in regional languages, and studies for which BurnedInAnnotation could not be confirmed NO.

Annotations

Reports were written by board-certified radiologists during routine clinical work and minimally edited for de-identification.

Phase labels in cect_3phase are derived from the DICOM SeriesDescription (e.g. "Arterial Phase Abdomen"). Series without a clear phase descriptor are labelled unknown.

The diagnosis column is reserved. Consumers can populate it from the report text using their own NLP pipeline.

De-identification

The release follows the DICOM PS3.15 Annex E Basic Application Confidentiality Profile. Identifying header tags were stripped from every DICOM (patient name, ID, birth date, accession number, institution and physician fields, station name, all private tags, verifying-observer tags). Acquisition dates were collapsed to year precision. BurnedInAnnotation was confirmed NO for every series. Reports were converted from PDF and DOCX to plain text and machine-redacted before release.

Subject IDs (PA{n}) are stable within this release and do not link back to source records. PatientSex and PatientAge (year precision) are retained. Audit artefacts are held by Perle AI and are not part of the public release.

Limitations

The cohort is small. With 10 subjects across three configurations, this is a development resource, not a benchmark.

All studies came from a single Siemens SOMATOM go.Now scanner under one institutional protocol set. Demographic, scanner, and disease-prevalence distributions reflect that source, not a global population.

All five cect_3phase subjects are male. gb_carcinoma is a single 43-year-old female. ct_general spans both sexes and ages 4–58. The sample is too small to support claims about sex- or age-related model behaviour.

One ct_general subject is a minor (4 years old). Pediatric uses require additional ethical review by the consumer.

Some ct_general series use filenames without a .dcm extension. Loaders should treat any file under */ST*/SE*/IM* as DICOM regardless of extension.

Ethical statement

The data was collected by Perle AI under informed consent from subjects, or from legal guardians for minors, for use of de-identified imaging in research. Data handling complies with applicable health-data-protection regulations (HIPAA, GDPR, India's DPDP Act, and equivalent local regimes). The release was reviewed under Perle AI's internal data-governance program.

Takedown requests can be raised through the dataset's Hugging Face discussion page or by emailing hello@perle.ai. Perle AI will respond within 7 business days.

How to use

from datasets import load_dataset

ds = load_dataset("Perle-ai/multimodal-ct-radiology-reports", "cect_3phase", split="train")
print(ds[0])

To load a single DICOM series as a numpy volume:

import pydicom, numpy as np, glob, os
slices = [pydicom.dcmread(f) for f in sorted(glob.glob(os.path.join(ds[0]["image_dir"], "IM*")))]
slices.sort(key=lambda s: float(s.ImagePositionPatient[2]))
volume = np.stack([s.pixel_array for s in slices])

Citation

@dataset{perle_ai_ct_radiology_2026,
  title     = {Perle AI Multi-phase CECT and CT with Radiology Reports},
  author    = {{Perle AI Research Team}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Perle-ai/multimodal-ct-radiology-reports}
}

Maintainers

Perle AI Research Team. Contact: hello@perle.ai. Maintainers: mad.elshiekh@perle.ai.

Changelog

See CHANGELOG.md.