In March 2009 I wrote a short piece on prefetching for new PACS administrators. The whole argument fit in four short paragraphs: an order arrives from the RIS, the PACS pulls prior studies from the archive to local cache, and the radiologist opens the study with the right comparisons already there. Spend time with your radiologists on what “relevant” means, or lose that time at the reading station.
I stand by the advice. I also wouldn’t write it that way now, because almost every noun in it means something different than it did. This is the first of what I expect will be several pieces revisiting the Archives with what I’ve learned since.
What hasn’t changed#
The purpose. A radiologist reading a chest CT needs the prior chest CT and probably the prior chest x-ray, and needs them before opening the study, not after a thirty-second wait. Report turnaround still depends on it. Hanging protocols still break without it. And the single most useful thing a PACS administrator can do about it is still a conversation with the radiologists about which priors matter for which exams. Software can apply relevance rules all day; it can’t decide what the rules should be for your radiologists.
The trigger, mostly. An inbound order still starts it: an HL7 v2 ORM, an OMI in newer interfaces, or, where FHIR is in play, a Subscription notification about a ServiceRequest. (ORM is HL7’s general order message, withdrawn in v2.7 but still what IHE Radiology’s HL7 v2.3.1 transactions use; OMI is the imaging order message HL7 defines for systems such as a RIS and a PACS.) Some systems also start from scheduling changes, status updates or worklist activity. Whatever the event, it carries or points to patient identity, the procedure, and sometimes a body part (an R4 Subscription notification may carry no payload, leaving the receiver to fetch the ServiceRequest), and the system uses those to decide what to look for.
What changed#
We still call it prefetching, but the word now covers several jobs. Sometimes it means moving DICOM objects ahead of time, the way it did in 2009. Just as often it means finding out which priors exist, deciding which ones matter, and telling a viewer where to get them. The clinical goal is the same; the work underneath it isn’t.
The archive stopped being a cache tier#
In 2009 “prefetch” meant moving studies from slow storage (tape, optical, a nearline array) to fast storage (the PACS’s local disk) so the viewer wouldn’t stall. The archive was deep and the cache was shallow, and the whole point was crossing that gap ahead of time.
Most of the enterprise archives I’ve worked on recently are flat: everything is on disk or object storage, and the difference between “online” and “nearline” is a few hundred milliseconds, not a tape mount. That removes most of the old storage-tier argument for prefetching. Most, not all. Fast storage isn’t the same as fast delivery: a two-thousand-image CT still has to cross the network, and a reading room two hundred miles from the data center feels every hop. And some archives I’ve seen push older studies to a cold cloud tier where a restore takes hours; for those, the 2009 logic applies unchanged.
So prefetching isn’t obsolete, but storage speed is no longer the main reason for it. The problem moved from storage latency to system boundaries.
The priors are in a different system#
In 2009 the PACS and its archive were usually one product, or at least one vendor’s problem. The prefetch was a PACS asking its own archive for its own studies. In the environments I see today there’s a vendor-neutral archive holding everything, a PACS (or two, after a merger) holding a working set, a separate cardiology system, an enterprise viewer that reads from the VNA directly, and a reading worklist that may belong to none of them.
The prefetch now crosses products, and it’s two operations, not one. First the reading system has to find the prior: a DICOM C-FIND, or a DICOMweb search (QIDO-RS), against the VNA. Then it has to get it: C-MOVE or C-GET, or a DICOMweb retrieve (WADO-RS). Either can work while the other fails. The query comes back empty because it went out under the wrong patient ID, while the study sits in the VNA. Or the query finds the study and the retrieve fails. With C-MOVE, the archive refuses a destination AE title it doesn’t know, or can’t open the separate association back to it (a firewall, a wrong port). With any retrieve method, an AE restriction or access policy can block it. And both can succeed while the viewer still never puts the prior in front of the radiologist.
All of those reach me as the same sentence: “my priors aren’t there.” The first question is which step failed: finding, getting, or showing.
Patient identity is the hard part now#
The 2009 piece assumed one patient ID. After a decade of health system mergers, most of the archives I work on hold studies for the same person under several issuers of patient ID. Issuer of Patient ID (0010,0021) is the DICOM attribute that names the assigning authority behind a Patient ID, and it’s optional (Type 3), so DICOM doesn’t oblige a sender to fill it in. IHE Radiology makes it a required query key for DICOM image queries [RAD-14] only under its Enterprise Identity Option, and for every QIDO-RS Query [RAD-129]. A prefetch keyed on today’s MRN finds nothing from the hospital that was acquired in 2017 unless something, usually an EMPI or the VNA’s own patient reconciliation, maps the identifiers.
This is the number-one reason I see relevant priors fail in 2026, and I see it constantly in outside studies: imported from a CD or an exchange network under the outside institution’s identifiers, never reconciled to the local MRN, and then invisible to every prefetch that follows. It has nothing to do with prefetch logic. The logic is fine. The patient on the order and the patient in the archive are the same human with two identities, and nobody told the prefetch engine.
The reverse problem is worse. Two hospitals can each have an MRN 12345 that belongs to two different people. An archive that matches on Patient ID without the issuer can’t tell them apart, and then the failure isn’t a missing prior; it’s someone else’s.
Fixing either one doesn’t always mean rewriting headers. Some environments relabel imported studies with the local MRN. Others keep the original identifiers and maintain a cross-reference, so a query for one identity finds studies filed under the other; some also record the other identities in the header’s Other Patient IDs Sequence (0010,1002); in the IHE queries named above it is a required return key, not a matching key. IHE’s Patient Identifier Cross-referencing (PIX) profile describes the cross-reference approach. In Radiology, Import Reconciliation Workflow (IRWF) covers reconciling imported outside studies to local identifiers at import, and Patient Information Reconciliation (PIR) propagates patient merges, ID-space mergers, demographic updates and corrections, including for images acquired under a temporary or wrong identity. Knowing which approach your archive takes tells you where to look when a prior goes missing.
Relevance got more interesting and less tractable#
In 2009 relevance meant same modality, same body part, maybe a lookup table that said chest CT also wants chest x-ray. That still covers most of it.
What’s new is that the body part is often wrong or missing in the header, procedure codes vary by site and by year, and the archive holds things few archives held in 2009: tomosynthesis stacks, enhanced multi-frame objects, dermatology photos, pathology slides, encapsulated PDFs of outside reports. “Relevant prior” for a breast MRI now plausibly includes a tomosynthesis study from another vendor, an ultrasound, and a scanned outside report. The lookup table needs to be a mapping layer, and the mapping layer needs to be maintained by someone who talks to the radiologists. Same advice as 2009; bigger table.
And finding everything isn’t the same as finding what’s useful. A patient with twenty prior abdomen CTs doesn’t need all twenty queued for the comparison pane. The most recent one and an older baseline usually matter more than the eighteen in between, though an oncology follow-up may want the whole run. Relevance is now a question of order as well as a yes or no.
The viewer moved#
In the products I’ve worked with, zero-footprint and server-side rendered viewers generally don’t keep a persistent diagnostic cache on the workstation. Depending on the product, they cache on a rendering tier, hold data in browser memory, or stream on demand and rely on the archive and the network being fast enough. Which brings us back to the first point. When the data is close and fast, the critical question is less “did we move the bytes” and more “does the viewer know the prior exists, recognize it as relevant, and load it in the comparison pane quickly enough to be useful.” Findable isn’t the same as ready: a large prior can show up in the study list and still take long enough to open that the radiologist reads without it. That’s a worklist and hanging-protocol integration problem, and it’s solved by the systems agreeing on identity, relevance and workflow. Fast storage alone solves none of those.
What I’d write today#
If I were giving the 2026 version of that advice to a new administrator:
- Find out where the priors physically are. VNA, legacy PACS, the cardiology system, the outside-image import box. If any of them are on tape or a cold cloud tier, the 2009 article applies to that one unchanged.
- Find out how many patient identifiers the archive knows about, and who reconciles them. If the answer is “nobody,” that’s your prefetch project, and it isn’t about prefetching.
- Trace one order end to end. Order arrives; what fires; what query goes to which archive with which patient ID, issuer and matching keys; what comes back; what retrieve follows, and to where; what the viewer does with it. Do this with a packet capture or the archive’s query log, not with the vendor’s architecture slide.
- Then have the relevance conversation with the radiologists. It is still the highest-value hour in the project, and it still comes after the plumbing is understood, not before.
- Check the result at the reading station. A prefetch job that reports success tells you a query ran and data moved. Open a handful of real exams the way a radiologist would, and see whether the right priors are there, in the right pane, when the study opens.
The 2009 article ends by telling administrators to put that time in with the radiologists up front, especially for the abdomen, pelvis and chest, where several modalities compete to be the relevant prior. That’s the part that aged best. The bytes are easier to reach now. Knowing which ones matter, whose they are, and getting them in front of the radiologist before the study opens is the hard part.
Sources#
- The original: Understanding the PACS Prefetching Process, JS PACS, March 2009.
- DICOM PS3.3, Section 10.15, Issuer of Patient ID Macro, for Issuer of Patient ID (0010,0021) as the assigning authority and its Type: https://dicom.nema.org/medical/dicom/current/output/chtml/part03/sect_10.15.html.
- DICOM PS3.3, Section C.7.1.1, Patient Module, for Patient ID and the Other Patient IDs Sequence (0010,1002): https://dicom.nema.org/medical/dicom/current/output/chtml/part03/sect_C.7.html.
- DICOM PS3.4, Annex C, Query/Retrieve Service Class, for C-FIND, C-MOVE and C-GET: https://dicom.nema.org/medical/dicom/current/output/chtml/part04/PS3.4.html.
- DICOM PS3.18, Web Services, for the Search transaction (QIDO-RS, Section 10.6): https://dicom.nema.org/medical/dicom/current/output/chtml/part18/sect_10.6.html, and the Retrieve transaction (WADO-RS, Section 10.4): https://dicom.nema.org/medical/dicom/current/output/chtml/part18/sect_10.4.html.
- IHE Radiology, Import Reconciliation Workflow (IRWF): https://wiki.ihe.net/index.php/Import_Reconciliation_Workflow.
- IHE Radiology, Patient Information Reconciliation (PIR): https://wiki.ihe.net/index.php/Patient_Information_Reconciliation.
- IHE Radiology Technical Framework, Volume 2 (Rev. 23.0), Section 4.14.4.1.2.1 and Table 4.14-3 (Enterprise Identity Option) and Section 4.129.4.1.2 (QIDO-RS Query [RAD-129]), for Issuer of Patient ID as a required query key and Other Patient IDs Sequence as a required return key: https://www.ihe.net/uploadedFiles/Documents/Radiology/IHE_RAD_TF_Vol2.pdf.
- HL7 v2.5.1, Chapter 4, Order Entry: Section 4.4.12 (OMI, Imaging Order Message): https://www.hl7.eu/HL7v2x/v251/std251/ch04.html. HL7 v2.7, Chapter 4, Section 4.4.1, for ORM “retained for backwards compatibility only as of v2.4 and withdrawn as of v2.7”: https://www.hl7.eu/HL7v2x/v27/std27/ch04.html.
- HL7 FHIR R4, Subscription (server push notifications when a matching resource is created or updated): https://hl7.org/fhir/R4/subscription.html.
- IHE IT Infrastructure, Patient Identifier Cross-referencing (PIX): https://wiki.ihe.net/index.php/Patient_Identifier_Cross-Referencing.
- Observations about flat archives, outside studies, merged identifiers, and where prefetch fails in practice are from my own work and stated as experience.
