CIMILab-DataOps

Problem

Medical AI experiments fail the reproducibility test far too often because data preprocessing is undocumented, dataset versions are untracked, and preprocessing code is tied to a specific researcher's machine. CIMILab-DataOps addresses the data side of reproducibility.

Approach

CIMILab-DataOps will be an internal toolkit for managing medical imaging datasets through the full lifecycle: download, version-pinning, preprocessing, split creation, augmentation, and export. It will provide:

• Declarative dataset configuration files (YAML-based)
• Version-pinned dataset snapshots with checksums
• Modular, reusable preprocessing steps (resampling, normalization, skull-stripping)
• Automatic provenance logging: every model training run records exactly which data version it used
• Integration hooks for CIMILab-Bench's evaluation pipeline

Status

CIMILab-DataOps is in the design phase, expected to launch alongside CIMILab-Bench in Q4 2026.