Opportunistic findings: where autonomous AI touches patients first.

A 58-year-old presents to the emergency department with acute abdominal pain. The attending orders a contrast-enhanced CT of the abdomen and pelvis. The radiologist reads the scan, confirms diverticulitis, dictates the report, and moves on. Buried in the same image stack: a 4mm hypodense lesion on the left adrenal gland. Probably benign. Possibly an adenoma. Possibly something else. The radiologist did not miss it through incompetence. The radiologist missed it because the clinical question was diverticulitis, the queue has 40 more studies, and the adrenal gland was not why the scan was ordered.

This is an incidental finding - a clinically relevant observation on an imaging study ordered for an unrelated reason. The literature calls them “incidentalomas” when they are masses and “opportunistic findings” when the framing shifts from accident to intent: what if we looked on purpose?

The scale of the problem is large and well-documented. A 2021 meta-analysis in Radiology found incidental findings on 10-30% of CT scans depending on the body region, with follow-up recommendation adherence below 50% in most health systems. The ACR publishes a white paper series on managing incidental findings precisely because the volume overwhelms existing clinical workflows. In the United States alone, roughly 90 million CT scans are performed annually. At the low end of the incidental-finding rate, that is 9 million scans per year with something worth noting that may not get noted.

The human bottleneck is structural. Radiologists read under time pressure, are compensated per study, and are held accountable for the clinical question on the order. The adrenal lesion on the diverticulitis scan is simultaneously everyone's problem and no one's job.

Most proposals for clinical AI start with the hardest possible deployment surface: primary care triage, emergency differential diagnosis, oncology treatment planning. These are high-stakes, open-ended, and deeply embedded in clinical workflows that took decades to form. Deploying autonomous AI into any of them requires solving every deployment problem simultaneously.

Opportunistic screening has a different geometry. Four properties make it unusually well-suited as a first clinical deployment surface.

The data already exists. The CT scan has been acquired. The patient has already been exposed to radiation, already been billed, already been positioned on the table. An opportunistic screen runs over data that is sitting in PACS (the picture archiving system) regardless of whether anyone reads it for secondary findings. There is no additional patient interaction, no additional imaging order, no additional consent beyond what was already given for the primary study.

The task is narrow and well-defined. Detecting a vertebral compression fracture on a chest CT, or coronary artery calcification on a non-gated scan, or hepatic steatosis on an abdominal study - these are specific, measurable classification tasks with established ground truth. They are the kind of problem where AI systems already outperform radiologists in controlled evaluations. The FDA has cleared over 700 radiology AI products as of 2026, most targeting exactly this kind of narrow classification.

The baseline is zero.If no one was going to look for the adrenal lesion anyway, the cost of an AI false negative is the status quo. This is a critical property for first deployments. When the alternative is “nobody checks,” the acceptable error rate for an automated screen is far more forgiving than when the alternative is an experienced specialist. You are adding signal to a process that currently produces none for these findings.

The clinician stays in the loop by default. An opportunistic finding flagged by AI still requires a radiologist to confirm and a referring physician to act. The AI is not replacing a clinical decision. It is surfacing information that would otherwise go unsurfaced. This makes the liability question simpler: the AI is an additional check, not a substitute for one.

A single monolithic model that reads an entire CT for every possible finding is the wrong architecture for this problem. The clinical literature defines incidental findings across dozens of organ systems, each with different prevalence rates, different clinical urgency, and different ACR management recommendations. A vertebral compression fracture on a chest CT and hepatic steatosis on an abdominal scan have almost nothing in common clinically. Forcing them through the same model conflates training objectives and degrades performance on both.

The better architecture is a swarm: a coordinated set of narrow-task models, each trained for a specific finding type on a specific anatomical region, orchestrated by a routing layer that knows which models to dispatch based on the scan protocol, the body region, and the clinical context. A chest CT triggers the coronary calcium model, the vertebral fracture model, and the pulmonary nodule tracker. An abdominal CT triggers the hepatic steatosis model, the adrenal mass classifier, and the renal lesion detector. Each model produces a structured output: finding present or absent, confidence score, anatomical coordinates, and a suggested follow-up action aligned with ACR incidental findings guidelines.

This is an engineering problem more than a research problem. The individual models already exist in various stages of FDA clearance and academic validation. What does not exist is the orchestration layer, the integration with radiology PACS and RIS (radiology information systems), the structured output format that maps to existing clinical reporting workflows, and the aggregation logic that combines findings from multiple models into a single actionable addendum. The research is done. The deployment engineering has barely started.

Everywhere that matters clinically. The swarm produces a draft addendum - a structured list of flagged findings with confidence scores and management suggestions. The radiologist reviews the addendum before it enters the medical record. If the radiologist agrees, the finding gets appended to the primary report. If the radiologist disagrees, it gets discarded and the disagreement is logged for model retraining. If the radiologist is uncertain, it gets escalated to a subspecialist workflow.

This is a genuine addition to the radiologist's capacity, not a challenge to it. A radiologist reading 40 studies in a shift does not have the cognitive bandwidth to conduct a systematic secondary review of every scan for every possible incidental finding. The AI handles the systematic sweep. The radiologist handles the clinical judgment. The combination catches findings that neither would catch alone under real-world production conditions.

The automation-bias concern is real. If the AI says “no findings,” does the radiologist stop looking? This is the same problem Tesla faces with Autopilot, and the solution is the same: the system must be designed so the human is reviewing AI output, not monitoring for AI failure. The draft addendum is a positive prompt (“here is what I found, do you agree?”) rather than a negative assurance (“everything is fine”). The radiologist is always reading something, never being told there is nothing to read.

This is the question that kills most opportunistic screening proposals before they reach a hospital administrator's desk.

No CPT code exists for “AI-assisted opportunistic screening of incidental findings on an already-acquired CT.” The scan was ordered and billed under the primary indication. The radiologist's professional fee covers the primary read. An additional AI-driven secondary screen falls into a reimbursement gap: the data is free (already acquired), the compute is cheap (cloud inference on a batch of DICOM images), but the clinical act of interpreting and acting on the findings has no payer pathway.

Three economic models could close this gap. The first is institutional risk reduction: a hospital that catches a vertebral compression fracture on a chest CT avoids a downstream osteoporotic hip fracture that costs $40,000-$60,000 in acute care and significantly increases mortality in patients over 65. Under a capitated or value-based care arrangement, catching incidental findings is directly cost-saving. The second is malpractice exposure: a radiologist who fails to report a visible finding on a scan, even when the scan was not ordered for that purpose, faces growing liability exposure as courts tighten the duty-to-report standard. An AI screen that surfaces findings the radiologist can then formally acknowledge or dismiss reduces this exposure measurably. The third is the one that will ultimately prevail: CMS creates a reimbursement code for AI-assisted opportunistic screening, the way it eventually created codes for every cost-saving prevention measure it initially refused to pay for.

Until the third model arrives, the first two are sufficient for early deployments in health systems that are self-insured, operate under capitation, or have aggressive risk-management programs. The total addressable market for opportunistic screening under current economics is smaller than the clinical need. It is large enough to build on.

Every component of the harness gets its first real test here, and each one surfaces problems that will recur across every future clinical deployment surface.

Clinical integration. The swarm needs bidirectional access to PACS - read access to pull DICOM images for inference, write access to push structured addenda back into the radiology workflow. It also needs metadata from RIS (the radiology information system) to know which scan protocol was run, which body region was imaged, and which models to dispatch. The integration point is not the EHR directly; it is the DICOM/HL7 layer that sits between the scanner and the reporting system. Different from an Epic integration, similar in complexity, and subject to the same 12-18 month deployment timeline per institution.

Liability architecture. If the AI flags a finding and the radiologist dismisses it, who is liable for the missed diagnosis? If the AI does not flag a finding that was visible on the scan, does the AI vendor carry exposure? These questions need contractual answers before a hospital will turn the system on. Current malpractice doctrine has no framework for this, and the answers will likely vary by state. A regulatory scaffolding that maps liability allocation across jurisdictions is the prerequisite.

Regulatory surface. Each narrow-task model in the swarm is likely a separate FDA Software as a Medical Device submission. The orchestration layer that routes models and aggregates findings may require its own clearance. The draft addendum that enters the medical record is a clinical document with legal standing. Navigating this across the 50-state patchwork of medical practice acts - particularly as states like Utah create new categories for AI clinical activity - is a regulatory engineering problem that no existing radiology AI company has solved at scale.

Data standards. The structured output of the swarm - finding type, anatomical location, confidence, suggested follow-up - needs a standard envelope. DICOM SR (Structured Reporting) is the closest existing standard, but it was designed for human-generated reports, not for machine-generated finding classifications. An AI-native structured output format that maps cleanly to DICOM SR while preserving the provenance, model version, and confidence metadata required for audit trails does not exist yet.

Trust infrastructure. A radiology department chair will not deploy an AI screening system because a vendor showed them a ROC curve. They will deploy it after seeing 6 months of outcomes data from a comparable institution, confirming that their malpractice carrier will not raise premiums, and verifying that the radiologists on staff do not revolt. Building this trust takes pilot deployments, published results, and time.

This is the deployment surface we are building toward at Superposition. Every piece of the harness that opportunistic screening forces us to build - PACS integration, liability contracts, SaMD submissions, structured output standards, pilot-site evidence - carries directly to the next clinical surface and the one after that. The strategic argument for sequencing radiology as our second base camp rests on this compounding property: solve the harness here, and half the harness is solved everywhere.