Clinical Trial Imaging Core Lab Software Problems: The 7 That Void Reads, and How to Avoid Them

Why does the assumption that a PACS or a viewer covers this keep happening?

Because the images are the visible part. A team looks at the requirement, sees DICOM storage and a diagnostic viewer, and concludes that the imaging infrastructure is most of the work with some workflow on top. Budget follows that shape, and the workflow is treated as configuration.

It is the wrong shape because a picture archiving and communication system and a core lab system have opposite design goals. A PACS exists to give every authorised clinician the fullest possible view of a patient, joining studies across time and modality. A core lab system exists to give one specific reader a deliberately constrained view of a de-identified subject in a defined order, then prove afterwards that the constraint held. Every default in a PACS works against you, and the features you actually need, which are assignment, blinding enforcement, criteria implementation and reconstructable audit, are not there at all.

The fix: scope from the imaging charter rather than from the images. Write down the read paradigm, the criteria, the adjudication trigger, the reader eligibility rules and the blinding constraints first, and treat storage and viewing as the substrate underneath them. Use imaging infrastructure where it earns its place, then build the blinding, assignment, criteria and audit layers deliberately. If a proposal leads with the viewer, it has the priorities inverted.

What goes wrong with image intake and de-identification?

Data arrives from hundreds of sites by DICOM push, portal upload, secure file transfer and physical media somebody still has to load. It is inconsistent by nature. Slice thickness differs from the imaging manual on one scanner, contrast timing drifts, a series is missing, and a subject imaged on one scanner at baseline is imaged on another at follow up, which quietly undermines comparison.

The failure that actually hurts is identifiers. Tag level anonymisation handles the header, and teams stop there. Patient identifiers also live in the file and folder structure, and they are burned into the pixel data on ultrasound captures, scanned requisition forms and secondary capture series, where no header rule will ever find them. One of those reaching a reader or a sponsor is a privacy incident, and it is discovered by someone else.

The second failure is over correction. A blanket strip of every non required field breaks the temporal and spatial relationships the read depends on, and you find that out when a reader cannot register a follow up against baseline.

The fix: two mechanisms, not one. Tag level rules that strip and replace identifiers while preserving the acquisition parameters, geometry and timing the read needs, plus pixel level detection of burned in text with a human confirmation queue for anything the detector is unsure about. Burned in text detection is one of the few places a model does clearly honest work here. Then run technical quality control against the charter automatically, comparing headers to the acquisition specification and flagging missing series, scanner changes between time points and out of specification parameters, so a technologist reviews exceptions rather than opening every study.

Why do the sponsor and site interfaces break after launch?

Three interfaces cause most of the post launch trouble, and each fails in its own way.

Site upload fails on people rather than protocol. A site uploads a study to the wrong subject, or uploads the same study twice under two identifiers, or sends a study for a time point that does not exist in the schedule. If the intake accepts it silently, you now have a duplicate in a blinded queue and a reader may read the same subject twice.

The feed to electronic data capture or randomisation fails on timing and on identity. Results are expected in a defined window, and a mapping change on the sponsor side, a subject identifier corrected at the site, or a time point renamed mid study will break the join without breaking the transfer. The transfer succeeds, the results land against nothing, and nobody notices until a data review.

Archive retrieval fails on expectations. A re read is requested for a closed study, the images are on the cheapest storage tier, and retrieval takes far longer than anyone told the sponsor.

The fix: validate intake against the expected visit schedule and reject rather than absorb anything that does not match, with a site facing message explaining why. Reconcile the outbound feed by count and by identity on a schedule, so a broken join raises an alert rather than waiting for a data review. Agree archive retrieval times with sponsors contractually, and test a restore before you need one.

What happens when reconstructability and validation are not covered?

Two years after a submission, a reviewer asks how a target lesion in the liver was measured at week 24. The honest answer has to include the exact image and series, the annotation geometry, the measured value, the reader, the timestamp, the charter version and the criteria version in force, the reader's certification status on that date, and whether the time point was later re read under a directive.

Systems that store annotations as overlays baked into a file cannot answer that. Systems that store a derived response without the measurements that produced it cannot re run the derivation. Systems where a charter amendment overwrote the previous version cannot say which rules governed a completed read. Each of those turns a one minute question into a forensic exercise across three teams and a sponsor who is now nervous.

Alongside that sits validation. A system producing endpoint data falls under 21 CFR Part 11, which brings electronic signature requirements, access controls and a qualification burden. Teams that treat validation as documentation written at the end discover that the design decisions validation depends on were made months earlier and are now expensive to change.

The fix: build an append only event record where annotations are objects carrying provenance rather than overlays, store derived responses separately from the measurements behind them, and keep charter and criteria as versioned artefacts referenced by every read session. Start the validation plan during design, not after, and let it shape the audit trail rather than describe it.

Should you build custom or contract the read from a provider?

If you are a sponsor with one imaging endpoint in one study, contract it. Calyx, Clario, Median Technologies and ICON Medical Imaging will deliver the endpoint with the regulatory familiarity and the reader network already assembled, and reproducing that for a single study is not a rational use of capital. The same holds when imaging is exploratory rather than registrational.

It is worth being explicit about why you cannot simply licence what they run. These are service organisations first. When you contract them you buy the read as a deliverable and the platform comes with it. The operating system you would need in order to run reads yourself is not what they sell, so a licensing conversation usually ends where it started.

Build when the read workflow is something you sell or something your science depends on. That means you are or are becoming a core lab, you run imaging endpoints across a portfolio and per study service fees have become a large recurring line, you have your own reader network and want control of assignment and certification, you work in a therapy area whose scoring system providers treat as bespoke, or you need imaging results to feed randomisation or an interim analysis on a timeline a service handoff cannot support.

How do hidden costs get into the quote?

Criteria implementations priced as configuration. RECIST 1.1, iRECIST, Lugano, RANO, PCWG3 and the non oncology scoring systems each carry a real body of rules with target and non target lesion handling, measurement constraints and response derivation, and each needs medical review rather than only code review. The second criteria set is not a fraction of the first.

The viewer. A measurement tool for one modality and a diagnostic grade multi planar viewer with registration and segmentation are different orders of work, and the requirement is often written in one line.

Modality breadth. PET with standardised uptake values, magnetic resonance imaging with multiple sequences and ophthalmic imaging each bring their own handling, and adding one late is not a small change.

Storage over time. Imaging volume compounds, and the year two bill is rarely modelled in the year one budget. Tiering, retention obligations that outlast the study and retrieval performance all need decisions early.

Validation, which is a workstream with its own calendar rather than a document.

From Digital Heroes delivery experience, a first release with multi channel intake, two mechanism de-identification, charter driven quality control, reader assignment, a blinded read workspace with measurement tooling, one criteria implementation and adjudication runs $120,000 to $240,000 over 16 to 22 weeks. A full platform adding further criteria, reader certification and discordance analytics, sponsor turnaround dashboards, storage tiering, re read directives and an electronic data capture feed runs $320,000 to $750,000 phased over 10 to 16 months.

What separates a build that works from one that fails here?

Blinding enforced at the query layer. A reader session should be physically unable to fetch what it must not see, with the constraint set expressed as test scripts that actively attempt each breach. Screen logic is not a control, it is a preference.

Attention to the leak paths that user interface design misses. Exports, error messages, notification emails, worklist counts that reveal another reader's progress, and support accounts with broad production access. Ask a prospective partner what a support engineer can see during an incident, because the answer separates people who have run this from people who have designed it.

Criteria and paradigm as configuration over a common measurement model, so a charter amendment mid study is a configuration change reviewed by your medical lead rather than a deployment window your read queue has to wait for.

Reader supply treated as the real constraint. Certification per criteria set as a first class record with refresher dates and retained evidence, assignment that respects conflicts and blinding history, and turnaround dashboards that age by study and time point so an operations manager can act this morning. Watch inter reader discordance as an operational signal, because a rising adjudication rate usually points at charter clarity or reader training rather than at the drug, and sponsors will ask.

Ownership settled before kickoff. You should own the repository, the image storage accounts, the infrastructure and the validation package. At Digital Heroes all of it is the client's from the first commit. Image retention obligations outlast software suppliers, so portability is not a negotiating point.