What to Require From an Imaging Core Lab Before the First Patient Is Scanned

Most sponsors choose an imaging provider under time pressure. The IND clears, the sites need to open, and imaging is one line on a long list. The decision gets made on price, a capability deck and a reference call.

The problem is that almost nothing in that process surfaces the things that go wrong later. Below are seven areas worth requiring specifics on before the first patient is scanned, and the questions that get you those specifics. None of this is exotic. All of it is answerable by a competent provider in writing.

1. Acquisition

Imaging endpoints assume the images are comparable. That assumption is made at the site, on equipment nobody in the sponsor organization has seen, by technologists following a protocol they received by email.

Require the acquisition parameters in writing before site activation: modality, slice thickness, contrast timing, field of view, reconstruction kernel. Then require the part most sponsors skip — what happens when a site deviates. Who notices, how quickly, and what is done about the study that was already acquired. A provider who checks acquisition only when a reader complains is checking too late.

Ask: how does a deviation reach you, and how long does it typically take from scan to notification?

2. Baseline

The most expensive imaging error in an oncology trial is usually the first one. A patient is screened, a local read finds measurable disease, they enroll and they are dosed. Months later central review disagrees.

If the disagreement concerns whether the patient had measurable disease at baseline, the cost is already fixed: a patient exposed to an investigational agent who could not have benefited, a response that may reverse on adjudication, a paid re-read, and efficacy data that no longer means what it appeared to.

This is structural rather than anyone’s failure. The screening read happens under time pressure, by a general radiologist, against inclusion criteria written by someone else. Central review happens later by design, because that is where the endpoint lives.

Require central confirmation of imaging-based inclusion criteria before the site commits, with a defined escalation path when local and central reads disagree. Then ask the question almost nobody asks: what is your local-versus-central agreement rate at baseline? If the provider does not measure it, that is still an answer.

3. Who Reads Your Studies

Credentialed is a word every imaging provider uses. Very few define it, and the ones that do not are relying on you not to ask.

Require five things in writing.

Qualification and registration, stated specifically rather than as a category. The qualification each reader actually holds and the jurisdiction they are registered to practice in. Be wary of any provider implying equivalence between one country’s certification and another’s — no such status exists, and a vendor who blurs it is telling you something about how they handle harder questions.

Qualification per study, not per résumé. Reader qualification should be granted for your protocol: the reading criteria in use, training completed against your imaging charter, and qualification cases read before any subject data is assessed. A radiologist qualified on RECIST 1.1 for one study should qualify again for the next.

What lands in the study file. For every reader: qualification held, registration and issuing council, training institution, years of subspecialty reporting experience, protocol-specific training record and qualification results, and a current CV. Assembled at study start and kept current, available to your monitors and to an inspector — not assembled after a finding.

Named accountability. You should know who read your study and be able to see the evidence they were qualified to. Reader identity and every measurement event recorded in the audit trail.

Subspecialty coverage confirmed against your protocol rather than advertised in advance. A provider listing nine therapeutic areas is telling you what they hope to staff, not what they can put on your study next month.

4. The Reads Themselves

Require the read design in the imaging charter and not in a conversation: how readers are assigned, how blinding is maintained, what triggers adjudication, and who adjudicates.

The adjudication trigger is the one to press on. It should be written down in advance with a defined threshold, not decided case by case once a discrepancy appears. A trigger agreed after the fact is not a trigger.

If your asset is CNS-active, add one more. Systemic response is assessed on CT under RECIST and intracranial response on brain MRI under a different convention — frequently by a different reader at a different vendor. The divergence is rarely an error. It is two conventions applied correctly by two people who never spoke, and it surfaces exactly when you are trying to characterize CNS benefit precisely. Ask who reconciles them, and when.

5. What Actually Arrives

In 2026, imaging results still arrive as a PDF on a large number of trials, and somebody retypes them into the EDC. Every transcription step has an error rate. Small per field, but multiplied across target lesions, non-target lesions, time points, patients and studies it becomes a standing source of queries that consumes data management time for the life of the trial. It also creates a reconciliation problem: when the database and the source report disagree, resolving it means going back to a signed document and interpreting it.

Require structured, eCRF-ready output: each target lesion as a discrete field with location, longest diameter, time point and reader; non-target lesions with classification; sum of diameters computed rather than transcribed; response category with the measurements that produced it attached; new lesion flags.

Ask a prospective provider how their results reach your EDC. The answer tells you a good deal about everything else.

6. The Audit Trail

The question is not whether a provider says they are 21 CFR Part 11 compliant. Everyone says that. The question is what is reconstructable eighteen months later.

Require specifics: which measurement events carry their own audit record, whether reader identity is recorded per event, whether a superseded measurement remains visible with its reason for change, and how an inspector would be given access.

Ask them to show you an audit trail for a closed study. A provider who cannot produce one quickly is telling you what an inspection will be like.

7. AI in the Read

Most imaging providers now use AI somewhere in the workflow. Few will volunteer where.

Require four answers. What the AI does — detection, measurement pre-population, quality checks, or something that reaches an endpoint. On what regulatory basis it is supplied. Whether the model is versioned, and what happens to your study if the version changes mid-trial. And who signs.

That last one is where vague answers cluster. Research use only, said without a subject, sounds as though the signed report carries the limitation. It should not. A provider should be able to tell you plainly which components are research-use and confirm that a credentialed radiologist reviews and signs every case. Be cautious of anything described as predicting response to therapy — that sits close to an efficacy claim and rarely survives contact with a quality function.

8. Exit

Ask before you sign, not when the relationship is ending. Who owns the imaging data and the derived measurements. In what format they are returned. How long the provider retains them and under what obligation. What happens to the audit trail.

Imaging data outlives most vendor relationships. The terms are far easier to agree while everyone is optimistic.

Ten Questions to Put in Your RFI

None of these requires a technical background to ask, and all of them are answerable in writing. A provider who answers eight of ten precisely and tells you plainly where the other two are weak is a better partner than one who answers all ten smoothly.

GenPhase.ai is a clinical trial imaging provider for oncology sponsors — central eligibility confirmation, exploratory endpoint reads and structured EDC delivery, with credentialed radiologist sign-off on ONIX AI, our read platform built for GCP and 21 CFR Part 11. We publish our own answers to the questions above. Send us the protocol and the enrollment timeline; our clinical team typically responds within one business day, and you will have a scoped plan and pricing within five business days.