How Shaip’s Expert Cardiac CT Annotation Accelerates Early Cardiac Amyloidosis Detection
A clinical AI research group partnered with Shaip to build an end-to-end cardiac CT annotation and model-training workflow, converting radiologist criteria for early cardiac amyloidosis into governed, production-grade labels and features for downstream ML.
Project Overview
A clinical AI research group focused on image‑based diagnostics for complex cardiology use cases, seeking repeatable expert‑guided labeling at scale.
The client aimed to detect early‑stage cardiac amyloidosis from CT scans—signals that are subtle and often missed. They partnered with Shaip to build an end‑to‑end annotation and model‑training workflow, converting specialist know‑how into consistent labels and features for downstream ML.
Key Stats
Modality
Cardiac CT; high‑volume, multi‑batch cohorts aligned to expert criteria
SME Collaboration
Radiologists + data scientists in closed‑loop review cycles
Deliverables
Clinically tagged image sets + versioned annotation protocol
Model Impact
99.8% validated accuracy on target condition classification
Governance
Privacy‑preserving workflows and documentation traceability
Challenges
- Translating subtle, early‑stage imaging cues into an operational taxonomy.
- Maintaining labeling consistency across large, multi‑batch cohorts.
- Synchronizing radiologist feedback with iterative model training cycles.
- Preserving privacy and documentation rigor throughout delivery.
Solution
Codified radiologist criteria for early amyloidosis into a practical labeling guide with acceptance thresholds, escalation paths, and evidence tags to capture rationale.
Executed a radiologist‑in‑the‑loop pipeline: trained annotators applied structured tags; senior reviewers adjudicated edge cases; final gold labels fed training.
Trained and validated classifiers in iterative sprints; tracked per‑revision metrics to quantify taxonomy improvements. Validated accuracy reached 99.8%.
Multi‑layer QC with duplication checks, drift monitoring, and discrepancy dashboards.
Privacy‑preserving processes; versioned protocol docs; traceability from case → tag → decision artifact.
Project Scope
| Track | What We Did | Output | QC Gates |
|---|---|---|---|
| Taxonomy | Converted expert criteria into label schema | Semi-automated tools + visual QC | Identity protection with signal preserved |
| Metadata De-ID | DICOM tag scrubbing | Rule-based removal + whitelist | No PHI leakage in headers |
| Verification | Reviewer audits | Checklists; sampling plans | Measurable PHI risk reduction |
| Governance | SOPs & training | Audit trails; access controls | Reproducibility & compliance |
The Outcome
- 99.8% validated accuracy for the target classification, enabling deployment‑ready research.
- Faster iteration by embedding specialist feedback directly into training cycles.
- Reusable playbooks for future, multi‑site cardiology AI initiatives.
Strategic Impact: Expert tacit knowledge was transformed into a scalable, governed pipeline—boosting detection performance while hardening compliance.
Shaip translated specialist insight into a production‑grade annotation and training workflow—raising accuracy while accelerating experiments.
— Head of Imaging AI, Healthcare Research Partner
