Biomedical imaging data platform

Unified ingestion and preprocessing platform for biomedical imaging at scale — sixteen multi-center datasets totaling 150K CT volumes, with automated annotation mapping across DICOM and NIfTI.

AstraZeneca · Tooling & Infrastructure

Problem

Multi-center medical imaging data is fundamentally heterogeneous. A single research question — say, “train a lung tumor segmentation model robust to acquisition variability” — touches data from many sources: different hospitals, different scanner vendors, different annotation conventions, different file formats (DICOM, NIfTI), different metadata schemas. Without a unified data layer, every new model or analysis re-implements ingestion and preprocessing from scratch, with subtle inconsistencies that compound across the team.

Approach

A unified data ingestion and preprocessing platform that standardizes:

• File format normalization — DICOM and NIfTI volumes brought to a consistent on-disk representation with preserved provenance.
• Annotation mapping — automatic mapping of thousands of annotation masks (segmentations, lesion bounding boxes, classifier labels) to the corresponding volumes, across diverse annotation file standards.
• Metadata harmonization — schema unification across vendor-specific DICOM tag conventions, with explicit handling of missing or inconsistent fields.
• Versioned snapshots — reproducible dataset definitions tied to experiments.

Result

Sixteen multi-center datasets ingested, totaling 150K CT volumes, with automated mapping of thousands of annotation masks. Adopted as the team’s standard data layer, reducing the time-to-experiment for new modeling work substantially.

Stack

Python, PyTorch, MONAI, ITK / SimpleITK for medical imaging I/O, custom metadata schema, internal annotation tooling.