In medical AI, there's a saying: a model is only as good as the data it learns from. As we enter 2024, much of our work over the past quarter has gone into something less glamorous than new architectures but far more important — a data pipeline that clinicians, hospitals and regulators can trust.
Step 1: De-identification
Every DICOM study that enters our environment passes through an automated de-identification stage that strips or replaces protected health information from file headers. Because identifying details can occasionally be "burned in" to the pixels themselves, we added an additional image-level check to detect and remove embedded text. Only after both checks pass does a study become available for research.
Step 2: Expert annotation
Machine learning needs labels. We built an annotation workflow where findings are marked in 3D by trained annotators and then reviewed by board-certified radiologists. Difficult or ambiguous cases are adjudicated by a second clinician. This multi-reader approach gives us higher-quality ground truth and helps us understand where even experts disagree.
Step 3: Quality control
We automatically screen for corrupted series, incorrect slice ordering, mismatched metadata and unusual acquisition parameters. Each dataset is versioned, so every model we train can be traced back to exactly the data it was trained and tested on.
Step 4: Diversity by design
A model trained on scans from one hospital and one scanner brand can stumble when it sees data from elsewhere. We now track the mix of scanner manufacturers, protocols, slice thicknesses and patient demographics in our datasets, and deliberately fill gaps to make our models robust.
The payoff
With this pipeline in place, retraining a model on an expanded dataset has gone from weeks of manual work to a largely automated process. More importantly, it positions us to work with academic medical centers — which is exactly where we're heading next.
Dinmar Labs Awarded Research Innovation Grant