How ApertureAI actually works — the workflow, end to end

The cleanest way to evaluate a radiology AI vendor is to ignore the demos and read the integration diagram. DICOM in, AI something, report out. Every vendor says the same three things. The differences are in the arrows — what touches the study between arrival and sign-off, what the radiologist actually sees on the workstation, and what gets recorded for the audit trail.

This post walks through the ApertureAI pipeline as it is wired today. Not the marketing version. The operational version — the one that runs in production, with the seams and the supervision model visible. If you are a radiologist evaluating us, an administrator trying to understand the integration, or a clinical informatics lead auditing our workflow, this is the post to read.

The pipeline at a glance

The ApertureAI pipeline has six stages. Each stage has a defined input, a defined output, and a defined record of what happened — captured automatically, not reconstructed after the fact.

1. DICOM receive. A study arrives from the modality, PACS, or a peer-site gateway. We accept standard DICOM C-STORE and C-MOVE, with TLS in transit. Every association is logged.
2. Pre-processing and triage. The study is parsed for modality, anatomy, priors, and acquisition metadata. A small modality-specific model assigns a triage category (stat / urgent / routine) and a confidence score. Critical findings — pneumothorax on chest CT, intracranial hemorrhage on head CT, pulmonary embolism on CTA — are flagged at this stage.
3. Draft generation. Modality-specific and anatomy-specific models produce a draft impression and a structured findings section. Each sentence in the draft is bound to the slice(s), measurement(s), and prior study reference(s) that informed it. This binding is the audit trail.
4. Radiologist review. The draft surfaces in a fast review surface — not a separate model box, not a separate report box. The radiologist edits, accepts, or overrides each field. Edits are captured with timestamps and actor identity. Critical flags force an explicit acknowledgment before sign-off.
5. Sign-off and ORU dispatch. On sign-off, an HL7 ORU^R01 message is dispatched back to the originating PACS / RIS / EHR via the configured outbound channel. An ACK is required; the dispatch is retried until ACK is received.
6. Audit record persistence. The complete record — DICOM arrival, triage decision, draft generation, every edit, sign-off, ORU, ACK — is persisted as a single signed record. It is reproducible, queryable, and exportable for CMS audits and medical-legal review.

What the radiologist actually sees

The fastest way to lose a radiologist's trust is to put the AI in a separate window from the report. We do not do that. ApertureAI produces a single review surface where the draft impression, the structured findings, the linked images, and the prior comparisons all live in one pane.

The radiologist works in one report, top to bottom. They can accept the draft as-is, edit a sentence, or reject the whole draft and dictate fresh. Every action is captured. Critical-finding flags cannot be dismissed silently — they require an explicit acknowledgment, and the acknowledgment is logged.

The design rule we follow: the radiologist is the author of record. The AI is an assistant. The interface reflects that — the radiologist is always in the primary editing position, and the AI's suggestions are always subordinate to their judgment.

What the administrator actually sees

From the operations side, the pipeline is observable end to end. A single console shows study arrival rate, triage latency, draft generation latency, review latency, ORU dispatch latency, and ACK rate. Each metric has a defined target and an alert threshold.

There is no implicit failure mode. If a study is received but the draft does not arrive within target, the case is flagged. If an ORU is dispatched and the ACK does not arrive, the dispatch retries and the operator is notified. If a critical-finding flag is acknowledged without a corresponding edit, the case is reviewed by the supervising radiologist before sign-off.

The system is green when the pipeline is healthy. The feed is active when studies are arriving within the expected window. These are two separate signals, and we report them separately — a quiet feed is not the same as a broken feed, and an idle monitor is not the same as a broken system.

What the audit trail records

Every report has a corresponding audit record. The audit record contains, at minimum:

• The DICOM arrival timestamp and source association
• The triage category assigned, the model version, and the confidence score
• Every sentence in the draft, bound to the slices and measurements that informed it
• Every prior study referenced by the draft
• Every edit made during review, with timestamp and actor
• Every critical-finding acknowledgment, with timestamp and actor
• The final signed report and the sign-off actor
• The ORU dispatch timestamp, the receiving system, and the ACK

The audit record is the asset. It is what survives a CMS audit, a malpractice deposition, and a value-based-care contract renegotiation. The draft is the visible surface. The audit record is what we are selling.

What we explicitly do not do

It is just as important to define the boundary. ApertureAI does not auto-finalize reports. It does not bypass the radiologist for any case, including routine cases. It does not generate a confident-sounding impression for a critical finding without forcing human acknowledgment. It does not dispatch an ORU without a signed report. It does not silently fail open.

If the pipeline is degraded — model latency exceeds target, ORU dispatch fails, ACK rate drops — the system surfaces the issue to the operator rather than continuing to operate without supervision. Fail-closed is the default; fail-open requires explicit operator action and is itself logged.

Why this is harder than a foundation model

A foundation model that drafts a chest CT report with high accuracy is technically impressive. A foundation model that drafts a chest CT report with high accuracy, with a per-field audit trail, integrated into the existing PACS / RIS / EHR workflow, with critical-finding supervision, with ORU dispatch and ACK, with full operator observability — that is a system. Building a system is harder than building a model.

This is the path we chose. It is slower to deploy than a departmental AI license. It requires more integration work, more governance work, and more operational work. We believe it is the only path that produces a defensible, procurement-grade radiology AI platform — and we believe the procurement decisions of 2027 and beyond will reward it.

If you are evaluating ApertureAI for your practice, your group, or your health system, the right question is not "what is the model's accuracy benchmark." The right question is "show me the integration diagram, show me the audit record, and show me what happens when the pipeline is degraded." We will show you all three.

— Aldo Ruffolo, DO, MBA
Founder, ApertureAI
Connect on LinkedIn