North 2022

### Overview Susan Kruger presents the PICTURES program of work, managed for the University of Dundee's Health Informatics Centre in collaboration with Public Health Scotland, the University of Edinburgh, Abertay, and industry partner Aridhia. Funded by a Medical Research Council grant, the project expands the capabilities of safe havens and trusted research environments (TREs) in Scotland to handle large-scale, routinely collected medical imaging data (specifically DICOM radiology files) for AI training and healthcare research while safeguarding patient confidentiality. The core outcome is the live Scottish Medical Imaging (SMI) service hosted within Scotland's National Safe Haven.

### Key Topics - **Medical Imaging Challenges**: In Scotland, over 3.5 million radiological examinations occur annually, generating massive quantities of unstructured DICOM data stored in the national PACS archive. While invaluable for training clinical AI models (e.g., detecting early signs of lung cancer, emphysema, or cardiovascular conditions), the data is unstructured, inconsistently labeled, complex, and filled with personally identifiable information (PII). - **The Five Safes Framework**: The infrastructure operates under the five safes model established in Scottish safe havens: safe people (accredited researchers), safe projects (governance ensuring ethical public benefit), safe settings (secure environments), safe data (de-identified minimal data), and safe outputs (statistical disclosure control prior to export). - **Governance and Legal Framework**: In Scotland, secondary use of pseudonymized healthcare data is legally permitted when public benefit is demonstrated and patient privacy is protected. Governance involves local Caldicott groups across 14 NHS health boards as well as the national Public Benefit and Privacy Panel (PBPP).

### Technical Details - **Data Ingestion and Scale**: In 2018, partner eDRIS (part of Public Health Scotland) obtained a research copy of over 1.5 petabytes of national PACS data spanning 2010 to 2018. - **Architecture and Security Zones**: SMI employs a microservices-based architecture conceptually divided into three audited zones: a restricted Red Zone housing raw identifiable data with automated processing, de-identified intermediate zones, and an analytic Green Zone. In the Green Zone, high-risk research is isolated inside Private Project Spaces (PPS) with automated software provisioning using Ansible playbooks (e.g., TensorFlow). - **De-identification and Redaction**: Patient identifiers reside in DICOM header tags, structured report text, and embedded image pixels. Initial releases use omission (single-layer primary images with minimal PII risk), while ongoing work employs tools like the Clinical Trials Processor (CTP) to redact specific pixel coordinates. Willful re-identification remains an offence. - **Cohort Building Tools**: To extract research-specific image IDs, the platform uses the Research Data Management Platform (RDMP) for cataloging, versioning, and querying DICOM metadata; SemEHR (Semantic Electronic Health Record) natural language processing for free-text structured reports; and autoencoders for unsupervised image representation, feature extraction, and similarity-based image retrieval.

### Takeaways The live SMI service enables population-scale, real-world imaging data to be safely interrogated for clinical AI development without compromising patient privacy. By automating extraction, standardization, and de-identification within a governed safe haven, the PICTURES platform facilitates reproducible medical research and helps alleviate radiologist workload.