Gains:
- Understanding how artificial intelligence is used in protocol recommendation, dose optimization, denoising and artifact detection, and the role of the technician.
- Ability to manage retake decisions and image quality flags with artificial intelligence support and reduce unnecessary dose and delay
- Understanding the risk that image enhancement algorithms may change the real anatomy (adding artificial detail) and the diagnostic validity limit
A good diagnosis in radiology starts with a good image. An image taken with the wrong protocol, blurred from motion, noisy, or with insufficient contrast can mislead even the best radiologist in the world. At the same time, imaging is a matter of dose: CT and x-ray use ionizing radiation, and unnecessary dose harms the patient. These two create a balance: if you lower the dose, the image becomes noisy; If you push the quality, the dose increases. Artificial intelligence is increasingly used in this balance — protocol recommendation, dose optimization, denoising, artifact detection and retake decision — and the main protagonist of this unit is often the x-ray technician.
But this field has its own, insidious risk, and it forms our core principle: While image enhancement algorithms reduce noise, they can add an artificial detail that does not exist in reality or erase a small real finding. The diagnostic validity of the enhanced image may be limited; In case of doubt, the raw/original image is returned. Looking good aesthetically does not mean being diagnostically accurate.
Before shooting: protocol and dosage
The protocol is a set of settings that determine how an examination will be taken: slice thickness, contrast use and phase, acquisition area, dose settings. Wrong protocol means wrong examination — for example, vascular pathology cannot be seen in an examination performed without contrast. AI can recommend the appropriate protocol according to the request and indication (medical reason that requires the examination), but the final decision lies with the technician and the radiologist.
Dose optimization: AI can recommend the lowest diagnostic dose according to the patient's body structure and examination purpose; The "ALARA" principle (As Low As Reasonably Achievable — the lowest reasonably achievable dose) is the fundamental philosophy of the field. AI-based reconstruction in modern scanners attempts to produce acceptable images even at low dose.
Stage
AI contribution
Who decides
critical risk
Protocol selection
Recommendation according to indication
Technician + radiologist
Wrong protocol = wrong examination
Dosage adjustment
Low diagnostic dose recommendation
Technician (ALARA)
Extremely low dose = diagnostic loss
Reconstruction/denoising
noise reduction
Auto + control
Artificial detail/loss of real findings
Artifact detection
movement/metal flag
technician
Missed artifact = misreading
Retake
quality flag
Technician + radiologist
Unnecessary dose vs missing test
Post shooting: noise reduction and artifact
After the image is captured, the AI does two typical jobs. Noise reduction (super resolution): "Cleans up" a low-dose, noisy image. This is a powerful tool because it can reduce dosage and maintain quality. But that's where the risk lies: as the algorithm fills in the noise by "predicting" it, it may add subtle structure that doesn't exist (artificial detail/hallucination) or delete a tiny real finding that resembles noise. Artifact detection: AI can flag quality issues such as motion blur, metal artifact, wrong position, and warn the technician about reshooting; This manages the balance between unnecessary dose and incomplete examination.
Caution: An enhanced image may be considered more reliable because it looks aesthetically "cleaner". However, denoising can erase a true micronodule or a thin linear fracture by confusing it with noise. In a questionable area, always return to the raw/original reconstruction and re-evaluate with the appropriate protocol if necessary.
Re-shooting decision: dose or diagnosis?
The technician constantly strikes a balance: if image quality is inadequate, re-acquisition saves the diagnosis but gives the patient an additional dose; If the patient considers the quality as "manageable", the dose will not be added, but the diagnosis will be at risk. The AI quality flag supports this decision: "this graph is not diagnostic of motion, it is recommended again". But the final decision lies with the human, because clinical urgency (for example, the risk of repositioning a major trauma patient) is a context unknown to the model.
three mini cases
Case 1 — Low dose gain. In a pediatric patient, AI-based reconstruction still produces a diagnostic abdominal CT, reducing the dose by approximately 40%. The radiologist confirms the image, and the diagnosis of appendicitis is made clearly. Here AI supported the ALARA principle; The radiologist confirmed diagnostic adequacy. The dose gain is real, but the image is diagnostic, the human confirms.
Case 2 — Artificial detail trap. On a low-dose chest CT, the denoising algorithm creates what appears to be a 3-mm “soft nodule” in the right lung. The radiologist becomes suspicious and returns to crude reconstruction; There is no real structure in that area — it is an artificial detail produced by the algorithm. The radiologist does not write the finding into the report. If the improved image were to be blindly trusted, the patient would enter an unnecessary chain of follow-up.
Case 3 — Escaped artifact. In an ICU portable radiograph, the AI does not flag (miss) the motion blur and the technician accepts the image. The radiologist is about to interpret the cloudy lower lung zone as "possible infiltration", but realizes that the quality is not diagnostic and requests it again. In the clear graph, the area is completely clear. Lesson: AI can miss artifact; Quality assessment is also done through human eyes.
Weak prompt / Strong prompt
Weak prompt:
You choose this BT protocol, do the best.
No indications, patient information and clinical questions; The model produces a blind recommendation and the responsibility for the decision is placed on it.
Powerful prompt:
Your role: ASSIST the technician and radiologist in protocol selection. Decision making; present options and justifications, leave the final decision to us. I will give you indications and patient information. List possible appropriate protocol options (contrast/no contrast, phase, section) with justification; Observe ALARA in terms of dosage; mark decision points such as "the radiologist must confirm the necessity of contrast for this examination." Indication: [write]. Patient: [age/gender/kidney function/allergy/pregnancy].
Strong claim clarifies indication, dosage and decision points; keeps the responsibility in the person.
Copiable prompt templates
PROTOCOL RECOMMENDATION TEMPLATEYour role: ASSISTANT in protocol selection. I will give you the indication and patient information. List appropriate protocol options (contrast, phase, section, field) with justification; Observe the ALARA dosing principle; Mark critical decision points such as contrast/pregnancy/kidney. The final decision lies with the technician and the radiologist. Information: [write]
DENOISING VALIDITY REMINDER TEMPLATEI will evaluate a low dose / enhanced image. Remind me of the risks of denoising/super resolution: adding artificial detail, deleting real small findings, aesthetic-diagnostic separation. Produce a checklist that reminds you to return to raw reconstruction in the questionable area.
RESHOOTING DECISION TEMPLATEI'll give you an image quality issue and clinical context. Conduct an evaluation weighing the diagnostic benefit of re-extraction versus the additional dose/risk; Also consider the risk of repositioning in an emergency/trauma context. The decision lies with the technician and the radiologist, you make suggestions. Context: [write]
ARTIFACT CONTROL TEMPLATEI will give you a modality and image description. List the possible types of artifacts in this image (motion, metal, position, noise) and how each might confuse the diagnosis; Remind the radiologist to confirm whether the quality is diagnostic. Information: [write]
Common mistakes
- Blindly trusting the improved image. Denoising can add artificial detail; When in doubt, one should return to the raw image.
- Delegating the protocol decision to the model. Contrast/phase/pregnancy decisions are clinical responsibility.
- Lowering overdose. Too low a dose impairs diagnostic adequacy; ALARA is required to be "diagnostic".
- Full trust in the AI artifact flag. The model may miss artifact; Quality should also be evaluated through human eyes.
- Applying reshooting mechanics. The risk of repositioning in an emergency/trauma context should not be underestimated.
Tip: When you see a small, borderline finding in an enhanced image, reflexively ask: “Is this present in the raw image?” Turning to raw reconstruction is the safest way to distinguish between the ghosts that denoising creates and the realities it erases.
In summary
Image quality, protocol and dose are the invisible basis of diagnosis. In this area, AI recommends protocols, optimizes dose (ALARA), reduces noise, flags artifact and supports the reshoot decision; It is especially powerful in producing diagnostic images at low doses. But image enhancement algorithms have an insidious risk: when filling in noise, they can add artificial detail or erase real subtleties. Although the enhanced image appears clearer, its diagnostic validity may be limited; In case of doubt, the raw/original image is returned. Since protocol, dose, and redraw decisions carry a clinical context, the technician and radiologist have the final say.
Application task
Select a study type from your routine. List protocol options and decision points (contrast, pregnancy, kidney) using the "Protocol Recommendation" template with an indication and sample patient information. Then create a low-dose/enhanced image scenario and write down what steps you would take in a suspicious area with the “Denoising Validation Reminder” list. Finally, discuss how you would decide on a reshoot dilemma (quality is poor but the patient is severely traumatized) with the “Reshoot Decision” template.
checklist
- [ ] I evaluated the protocol according to the indication; Contrast/pregnancy/kidney decisions were made by humans.
- [ ] I adjusted the dose according to the ALARA principle, maintaining diagnostic adequacy.
- [ ] I confirmed the suspicious finding on the enhanced image with raw reconstruction.
- [ ] I took into account that denoising may add artificial detail/delete the real finding.
- [ ] I evaluated the artifact both with the AI flag and with my own eyes.
- [ ] I took the clinical context (urgency, patient condition) into account in the decision to reshoot.
- [ ] I did not confuse aesthetic appearance with diagnostic accuracy.