Healthcare AI2026 / 03

Radiology Value Pipeline

A radiology workflow platform that connects imaging models to human review, audit trails, governance, and Power BI reporting, with its MRI evidence clearly bounded.

FastAPIPostgreSQLClinical AIPower BI

Source & deliverables

Radiology review workspace showing a brain MRI

20public CUDA MRI volumes

120Kdeterministic evaluations

123,020engineering evaluations

0.9828held-out MRI benchmark AUC

01 / Context

The question
behind the work.

Radiology AI is often presented as a model score, while the harder operational questions remain unanswered: who reviews the output, how versions are traced, what happens after an override, and how quality and governance are measured over time.

My role

I designed and built the complete MVP: product specification, FastAPI backend, PostgreSQL schema, worker orchestration, imaging-model paths, review workspace, audit and feedback systems, SQL marts, Power BI package, tests, and documentation.

02 / Implementation

What I built.

The platform distinguishes 20 real public CUDA MRI volumes from 120,000 deterministic engineering evaluations and retains study, model, prompt, findings, feedback, and reporting lineage for review.

Registered chest X-ray and brain MRI studies through a normalized ingestion contract with demo and DICOMweb-oriented adapters.
Orchestrated ordered specialist services for representation, view quality, abnormality detection, urgency, explanation, report drafting, and discrepancy review.
Persisted model runs, model versions, prompt versions, structured findings, report drafts, clinician feedback, final reports, audit events, and governance snapshots in PostgreSQL.
Built a browser review workspace and three SQL-first reporting marts for operations, clinical quality, and governance.
Added real-versus-stub labeling so the platform never implies that every component is clinically validated.

03 / In detail

A workflow around evidence

I designed Radiology Value Pipeline around the work that follows an imaging prediction. A study enters through a normalized ingestion contract, moves through an ordered pipeline, and leaves behind structured findings, urgency suggestions, report drafts, feedback events, final reports, discrepancy records, and audit history.

FastAPI handles the product APIs, PostgreSQL holds the operational state and lineage, and a database-backed worker claims jobs with row locking. The modular-monolith design keeps the system runnable and understandable while preserving clean boundaries for ingestion, pipeline services, reporting, feedback, governance, and dashboard marts.

MRI evidence boundary

The MRI evaluation includes 20 real public CUDA volumes from the public MSD Task01 BrainTumour dataset. Those runs cover prediction conversion, scoring, review overlays, and candidate cards. The 120,000 deterministic robustness, acceptance, conversion, and contract evaluations are separate engineering checks across synthetic failure modes and workflow gates; together the tracked engineering total is 123,020 case-evaluations.

The public MRI cohort ran with a 240 x 240 x 160 ROI. It completed with mean whole-tumor Dice of 0.9202, mean region macro Dice of 0.8411, mean whole-tumor HD95 of 7.3455 mm, and mean whole-tumor 2 mm surface Dice of 0.2970. The 2D held-out benchmark macro AUC was 0.9828. Every candidate card remains reject_for_more_evidence.

The product surface

The browser workspace gives a reviewer the study context, image preview, candidate findings, model posture, evidence references, draft report, feedback controls, discrepancy events, agent ledger, and audit trail in one review surface. Structured findings come before the human-readable report text, which keeps the information usable downstream.

I also built SQL-first marts and Power BI outputs for operations, clinical quality, governance, model inventory, and agent lineage. The portfolio run processed 76 studies, stored 129 candidate findings, recorded 304 workflow-agent runs, and produced 858 dashboard fact rows.

04 / Engineering judgment

The decisions
that shaped it.

Used a modular monolith plus a database-backed worker instead of premature microservices. Row locking with SKIP LOCKED provides safe concurrent claims without adding a queue broker to an MVP.
Kept DICOM binaries and generated artifacts outside PostgreSQL while storing normalized metadata, retrieval references, and complete lineage in the operational database.
Kept clinicians in the review loop through structured feedback, report drafting, final sign-off, and complete audit history.
Kept Power BI downstream of explicit SQL marts so dashboard logic remains inspectable and reproducible.

Evaluation & results

The MRI evidence is deliberately separated: 20 real public CUDA MRI volumes completed with a mean whole-tumor Dice of 0.9202, while 120,000 deterministic robustness, acceptance, conversion, and contract evaluations verify engineering behavior. Together the tracked engineering total is 123,020 case-evaluations. The held-out MRI benchmark macro AUC was 0.9828; every candidate remains reject_for_more_evidence.

05 / Working outputs