Gains:
- Understanding how the PACS, RIS, DICOM and worklist chain works and where the AI output (marking, measurement, priority) is added to this flow
- Ability to understand how artificial intelligence results are displayed as a separate series/marking, that the original image should never be modified, and the importance of the audit trail
- Integration security, ability to design safe-default behavior in case of incorrect patient match and system outage
In a radiology unit, the image passes through an invisible pipeline from "the moment it is captured" to "the moment it is reported." Connecting AI to this flow without understanding this pipeline is like punching a hole in the wall without seeing the electrical wiring: connect it in the wrong place and you could corrupt the original image, add results to the wrong patient, or never consider what will happen if the system crashes. In this unit, we will explain the chain — RIS, PACS, DICOM, worklist — and show where the AI output will be added to this flow and with what security rules.
The core principle is that the AI output adds to the original diagnostic image, not replaces it. The original DICOM image is never modified; AI marking, measurement and prioritization are stored as a separate series/layer and each step is written to the audit trail. The radiologist should be able to see the raw data at all times.
Let's get to know the chain
- RIS (Radiology Information System): System that manages examination orders, appointments, patient demographics and report workflow. The "who, when, which examination" information stops here.
- PACS (image archiving and communication system): System in which images are stored, displayed and distributed. The radiologist reads the image via PACS.
- DICOM (Digital Imaging and Communications in Medicine): Standard file and communication format of medical images. A DICOM file carries both image pixels and a header — such as patient name, ID, exam type, date, device.
- Worklist: List of examinations that the radiologist will read; It is fed by RIS and PACS.
- Modality: The device that produces the image (CT, MRI, x-ray, ultrasound). The image flows from the modality to the PACS as DICOM.
The flow is typically as follows: Clinician requests examination in RIS → patient is withdrawn, modality DICOM sends image to PACS → examination falls into worklist → radiologist reads in PACS → writes report to RIS → report reaches clinician. AI enters this flow at three typical points.
AI entry point
What does
How to store
security rule
After shooting (triage)
Generates priority score
Worklist flag
Does not eliminate the examination, changes the order
During reading (CAD/measurement)
Marking, measurement
Separate serial/layer/SR object
Does not replace the original
During reporting
draft text
RIS draft space
Radiologist verifies and signs
In the DICOM world there are special objects for AI output: for example DICOM SR (Structured Report) measurements; segmentation object lesion boundaries; It carries separate series markings. Common point: these are all added next to the original image, not overwritten.
Why the original is never replaced
There are three reasons. First, diagnostic integrity: The radiologist should be able to see the area not marked by the AI and access the raw data; This will be lost if the marking is permanently overwritten on the original. Second, distinguishability: The marking should be on a separate layer and can be turned on and off, so that the AI output does not interfere with the real anatomy. Third, audit trail: Who produced what output, when, with which AI version — this log is essential for bug inspection, accountability, and quality. When a false negative is reviewed later, the answer to the question "what did the model see that day, what did the radiologist confirm" is in the trace record.
Caution: Permanently “burn-in” the AI markup onto pixels of the original diagnostic image irreversibly contaminates the diagnostic data. The markup should always be a separate layer/series and the original image should be preserved unchanged.
Integration security: three critical risks
Incorrect patient match: If the AI output is linked to the wrong exam/patient, one patient's finding may end up in another's report. Matching should always be done with strong IDs (accession number - study access number, study instance UID - study unique ID); Name similarity should not be relied upon.
System outage and safe-default: What happens if the AI service crashes? The correct design works "fail-safe": If the AI output does not arrive, the exam still falls into the worklist and the radiologist reads it in the normal order. AI crashing should never result in a study being lost or unread.
Version and change management: When the model is updated, its behavior may change; it should be monitored which version produces which output, local verification should be done before updating (this connects with Units 8 and 10).
three mini cases
Case 1 — Separate series saves. A CAD system marks three microcalcifications on a mammogram. Marks are added as a separate layer; The radiologist closes the layer and examines the raw image, evaluating two of the marks as real and one as artifact. Because the original image remained unchanged, the radiologist was able to make his or her own independent reading. Because the layer was on/off, the AI output did not interfere with the real anatomy.
Case 2 — Wrong match caught. One day, there are two "Mehmet Yıldız" in the worklist. A weak rule of integration that matches by name rather than accession number is about to link one patient's triage flag to another's exam. Quality control realizes that the matching should be done with the study instance UID and corrects the rule. Matching the name could have been a patient safety accident.
Case 3 — Business doesn't stop when AI crashes. One night, the AI triage service crashes during an update. Thanks to the secure-default design, the exams still arrive in the PACS and the worklist operates normally (arrival order); The radiologist continues reading without interruption. The shuttle comes back in the morning. In the absence of AI, no examinations were lost because the system gave a "normal list" rather than an "empty list".
Weak prompt / Strong prompt
Here the AI text assistant is used when designing the integration rule or safe-default scenario (not image processing).
Weak prompt:
We will connect AI to PACS, tell us briefly how to do it.
No context, no security requirements; A general and risky answer comes.
Powerful prompt:
Your role: ASSOCIATE consultant to the radiology IT/integration team. Decision making; produce checklist and risk warning. We connect a triage AI to the PACS/worklist flow. Itemize security requirements for: (1) what strong identity should be used for patient/exam matching, (2) how the AI output should be kept separate from the original image, (3) what the safe-default behavior should be if the AI service crashes, (4) what fields should be kept in the trace log. Add "why" and "risk if done wrong" for each item.
The strong will role clarifies the security dimensions and rationale.
Copiable prompt templates
INTEGRATION SECURITY CHECK TEMPLATEYour role: integration consultant (assistant). When adding an AI printout to a PACS/RIS flow, generate a checklist on the following topics: patient/exam match ID, preservation of original image, separate batch/layer storage, secure-default behavior, track record fields, version management. Add risk note for each item.Context: [write]
SAFE-DEFAULT SCENARIO TEMPLATEI will give you an AI failure scenario (service crashed, delayed, failed). For each scenario, define the safe-default behavior of the system: will the exam still be read, how will it be reported to the radiologist, will there be data loss. Principle: AI failure should never result in a study being lost/unread. Script: [write]
MATCHING VERIFICATION TEMPLATEI will give you an audit match with an AI output. Evaluate whether the matching is made with strong IDs (accession number, study instance UID) or weak IDs (name, date of birth) and warn of the risk of mismatching. Context: [write]
TRACE LOG (AUDIT) DRAFT TEMPLATEProduce a draft of the trace record fields that should be maintained for an AI-assisted reading event: exam ID, AI model and version, output type and score, timestamp, radiologist ID, radiologist's approval/rejection decision. Purpose: post-review and accountability. Context: [write]
Common mistakes
- Permanently writing the AI markup to the original image. Diagnostic data is contaminated; The marking should be a separate layer.
- Matching patient to name. Patients with the same name are confused; Strong identity (UID/accession) should be used.
- Not thinking about safe-default. If the analysis is lost when the AI crashes, patient safety is compromised.
- Not keeping track records. Who-what-when recording is essential for error investigation and accountability.
- Not monitoring the model version. Update changes behavior; It should be known which version produces what.
Tip: As you design the integration, ask one question at each step: “Would this audit be read safely if AI did not exist?” The answer should always be "yes". AI is an accelerator; The safe operation of the system should not be dependent on it.
In summary
The radiological image flows in a chain consisting of RIS, modality, PACS, worklist and DICOM format. AI enters this flow at three points: post-shot triage (worklist flag), CAD/measurement during read (separate batch/layer), draft during reporting. The golden rule is that the original diagnostic image should never be altered; The AI output is stored as a separate layer/series/SR object and each step is written to the trace register. Integration security has three pillars: patient matching with strong identity, AI failure-safe-default behavior, and version management. AI is an accelerator; The safe functioning of the system should not be dependent on it.
Application task
Draw the image flow of your unit (or a sample organization): RIS → modality → PACS → worklist → report. Mark where a triage AI is added to this flow. Then, using the “Secure-Default Scenario” template, write what the system should do for three failure cases (service down, late response, bad output). Finally, list the fields you will keep in an AI reading event with the “Trace Log Draft” template.
checklist
- [ ] I have ensured that the original DICOM image has never been modified.
- [ ] I stored the AI output as separate serial/layer/SR object.
- [ ] I made patient/test matching with strong ID (UID/accession).
- [ ] I defined safe-default (scan still read) behavior on AI failure.
- [ ] I keep the model version, score, time, and radiologist decision in the trace record.
- [ ] I'm tracking model version changes.
- [ ] "Would it be read safely even if there was no AI?" I can say yes to the question.