Radiology and Imaging Center Software Problems: The 7 That Cost Real Money, and How to Avoid Them

Why does the scope keep getting written as a portal for referrers?

Nearly every imaging group that calls us opens with the same request: build us a referrer portal. It is an understandable instinct, because referrers are the customer and the portal is the visible thing. It is also the scope that fails most reliably here, and the reason is adoption. Referrers already live inside their own electronic health record, so asking them to log into a fourth system to read one report is a request most practices quietly decline, which leaves your staff faxing anyway.

The work that actually pays is upstream and downstream of the portal. Upstream is intake, where fax, emailed portable document format orders and phone calls become structured orders, and where the wrong procedure code, the wrong laterality and the missing clinical indication enter your system. Missing indication is the one that costs, because it drives the authorisation denial weeks later, after the scan has been performed. Downstream is delivery routing, where one finalised report fans out on the channel each specific provider actually wants, with delivery receipts.

Scope those two first and the portal becomes a thin, optional surface rather than the project. The version that does get adopted has no login wall past a link sent directly to the ordering provider, shows the report and key images, and offers one click to order the follow up, because the follow up order is revenue you currently lose to whoever is easier to order from.

What goes wrong when you migrate patient identities and legacy studies?

Two migrations run in parallel here and the identity one is harder than the pixel one.

Your picture archiving and communication system, or PACS, deduplicates on a medical record number, and that number is usually site specific. So the same patient exists two or three times across your estate, plus once more under a maiden name, plus once more from the centre you acquired that ran its own numbering. Merging those is not a lookup, it is a probabilistic matching exercise across name, date of birth, sex, phone and address, with a confidence band and a queue for the records that land in the grey zone. Groups that treat it as a deterministic join produce false merges, which are far worse than missed ones, because a false merge puts one patient's study under another patient's name.

The study migration is where budgets disappear. Moving years of studies out of a legacy archive is its own project with its own timeline, and it usually delivers nothing your referrers can feel. The cheaper path is a read only query and retrieve bridge behind your matching layer, which gets you the priors within a sprint or two and defers the full migration decision until the new platform is carrying live volume.

Sequence it as identity first, bridge second, migration last or never. And build the adjudication queue as a real piece of the product with a named owner, because an unowned queue fills up and stops being looked at within a month.

Why do the interface and imaging integrations break after launch?

Three connections carry the operational risk. Health Level Seven, or HL7, messaging with referrer systems for orders in and results out. Digital Imaging and Communications in Medicine, or DICOM, query and retrieve against your archives. And the reporting system, commonly PowerScribe, that signals when a report is final.

The interfaces break for a reason specific to this market: you do not control either end of most of them. A referrer's information technology team upgrades their electronic health record, a segment changes shape, and your inbound orders stop parsing. Nothing pages anyone, because a message that fails validation looks like an absence of messages rather than an error. Then a practice calls a week later asking why nothing has been scheduled.

The design that survives this is a per endpoint liveness monitor with an expected volume profile. If a referrer who normally sends a predictable daily flow sends nothing by mid morning, that raises an alert naming the endpoint, rather than waiting for a human to notice an absence. Hold every unparsed message in a visible queue with the raw payload attached so someone can see what changed.

What happens when consent, audit and subprocessor obligations are not covered?

Three obligations get treated as paperwork and all three are architecture.

The first is the business associate chain. If a model reads your faxed requisitions, that provider is handling protected health information and needs to be inside the chain in writing. Groups discover this at security review, after the intake pipeline is built, and then have to rearchitect around a provider that will sign. Settle the list of every subprocessor that touches patient data before design, not after.

The second is audit logging at the study level. The question that eventually gets asked is not whether your system was secure in general but who viewed a specific patient's study and when. That means per user logging on every view, not just on login, and it means role based access down to the site level so a scheduler at one location cannot browse another location's worklist without a reason.

The third is the extraction audit trail. When a model pulls demographics, procedure and indication from a faxed order, you need a permanent record of what was extracted, what confidence it carried, which human confirmed it and what they changed. That record is what lets you defend the process, tune the threshold and answer a question about a specific order months later. Build it with the pipeline, because reconstructing it afterwards is impossible.

Should you build custom or configure what your PACS and RIS vendors already sell?

Plenty of imaging centres should configure rather than build, and we say so on calls. If you are single site, running a moderate annual study volume with a small active referrer base and no acquisition pipeline, your PACS vendor's portal, the scheduling module in your radiology information system, or RIS, and a good fax server will serve you. The money is better spent on a second technologist. The same applies if your entire referrer base is one hospital system on one electronic health record: build that single interface, take the win, and stop there.

Sectra, Merge and Fuji Synapse are serious archives and PowerScribe is a serious reporting product. None of them are the problem. The honest constraint is that they all start at their own boundary. The archive begins at the imaging boundary and has no opinion about the faxes. The reporting system ends when the report is signed and has no opinion about how it reaches fourteen different providers on the channels they each prefer. The identity problem spans systems none of them own, which is why no vendor will solve it for you.

Build when three signals appear together: two or more sites without a shared patient index so someone reconciles identities by hand, more than one full time person whose actual job is retyping orders and faxing results, and an acquisition pipeline that keeps adding another archive to integrate. At that point the off the shelf stack is not saving money, it is converting your software budget into headcount and hiding it in operations.

How do hidden costs get into an imaging software quote?

Interface count is the first and it is the one that moves schedules rather than just budgets. Every referrer system interface has its own quirks and its own information technology department that answers in weeks, and that waiting is not something your developer can compress. Sequence intake and document extraction to go live before the interfaces land, so the project delivers value while you wait.

Legacy archive dialects are the second. An archive from an acquired site may speak a version of query and retrieve that predates anything current, and discovering that during implementation is expensive. Ask for a connectivity test against every archive before the contract is signed, not after.

Historic data volume is the third and it is the cheapest lever you have. Migrating years of studies is a project. Bridging to them is a sprint. Decide deliberately rather than defaulting to migration because it feels tidier.

Then the ones outside engineering: security review across every subprocessor, payer rules maintenance your authorisation team should own directly, protocol duration modelling per magnet per site, and the correction feedback loop on document extraction, which needs someone reviewing exceptions for several weeks before auto accept thresholds can be trusted.

What separates an imaging build that works from one that fails?

The builds that work make every failure visible instead of silent. That is the through line in this category. A failed disc import, an unmatched prior, an unparsed order message and an undelivered report should all surface as a state on a record with a reason attached, in front of a person whose job it is, before the radiologist or the referrer discovers it. Make each prior fetch a first class record with requested, retrieved, matched or failed states and a reason, and surface anything unresolved a day before the appointment rather than at read time.

The second marker is that a human stays in the loop where the model is uncertain. Faxed requisitions are the right shape for extraction because the input is semi structured, the failure is visible and a coordinator is already there. Route anything below your confidence threshold to a queue with the fax image beside the extracted fields so confirming takes seconds instead of keying taking minutes, and feed corrections back so the threshold can move over time.

The third is that delivery preferences live per provider, not per practice. One report finalises and fans out on the channels each named provider actually wants, with receipts, and critical findings get their own escalation ladder that does not stop until acknowledged, with a timestamped trail that survives a discovery request.

Finally, own the code, the repository and the infrastructure accounts from day one, and get the business associate chain in writing covering every subprocessor including whichever model provider touches your intake documents. A vendor who will not sign that or will not hand over the repository is selling a subscription with extra steps.