databricks-industry-solutions/pixels

What it solves

Pixels is a solution accelerator designed to handle the complexity of medical imaging data (DICOM files). It solves the problem of ingesting, indexing, and analyzing millions of medical images at scale while maintaining strict healthcare privacy standards (PHI redaction) and providing a seamless bridge between raw image data and AI-driven clinical insights.

How it works

The system uses a Lakehouse architecture to process DICOM files through several stages:

1. Ingestion & Indexing: It catalogs DICOM files from cloud storage (S3, ADLS, GCS) using Spark and Unity Catalog. It can handle zip archives and stream new files incrementally using Auto Loader.
2. Metadata Extraction: It extracts DICOM metadata tags into Databricks tables, allowing users to run SQL queries and build BI dashboards over the image metadata.
3. Privacy Protection: It applies format-preserving encryption to redact Protected Health Information (PHI) from both metadata and pixel-level data.
4. Visualization & AI: It integrates the OHIF Viewer for image viewing and labeling. It leverages NVIDIA's MONAI framework for automatic image segmentation and allows users to launch model training or real-time inference via serverless GPU clusters.
5. Overlay Support: It supports loading NIfTI (.nii.gz) segmentation masks as overlays on top of DICOM volumes.

Who it’s for

This project is for healthcare data engineers, medical AI researchers, and clinical analysts who need to manage large-scale medical imaging datasets and build AI segmentation models within a secure, governed environment.

Highlights

• SQL-powered Medical Imaging: Turns unstructured DICOM files into queryable SQL tables for cohort building.
• Integrated AI Workflow: One-button launch for model training and real-time segmentation using MONAI.
• Enterprise Governance: End-to-end security and access control via Databricks Unity Catalog.
• Privacy-First: Built-in metadata anonymization and PHI redaction tools.
• Scalable Ingestion: Supports streaming ingestion and automatic unzip of medical image archives.