Gains:
- Understand how artificial intelligence saves time in automatic registration with previous examinations, change detection and lesion tracking.
- Ability to validate automated measurements and make consistent follow-up decisions in response evaluation frameworks such as RECIST
- Being able to recognize mismatches and different protocol pitfalls of automatic comparison and make the final evaluation as a radiologist
In radiology, an image alone is often half a sentence; Its meaning is completed when compared with previous examinations. If a lung nodule is 8 mm today, is that good or bad? The answer depends on whether it was 8 mm (stable, probably benign) or 4 mm (growing, suspicious) six months ago. Whether the treatment works for a cancer patient can only be understood by comparing the lesions in the previous and current examination. Comparative reading and follow-up is one of the most time-consuming but valuable tasks in radiology, and artificial intelligence is a powerful assistant here: it saves minutes by automatically matching (registration) with previous examinations, detecting changes and tracking lesions.
Core principle: Automatic comparison speeds matching to previous examination and measurement of change; But there are pitfalls of mismatch, different protocols and method differences. The final assessment—whether the change is real or method/protocol-induced, decision on response to treatment—belongs to the radiologist.
How automatic comparison works
Registration (image alignment): It is the process of placing the examination on two different dates into the same spatial frame; thus a lesion can be shown in the same location on both images. When the patient is filmed in a different position, holds his breath, or is filmed on a different device, alignment becomes difficult and there is a risk of mismatching.
Typical flow:
- The previous exam is found and matched. The system retrieves the patient's old examinations and aligns them with the new examination.
- The lesions are paired. The model maps a lesion to its counterpart in a previous exam.
- The change is calculated. Diameter/volume difference, percentage change, growth rate are produced.
- Change is visualized. By side-by-side or overlapping (fusion) display, “emerging” lesions are marked.
- The radiologist confirms. It checks that the pairing is correct, the protocols are comparable, and the exchange is real.
RECIST and response evaluation
In oncological follow-up, RECIST (Response Evaluation Criteria In Solid Tumors) is a framework used to measure response to treatment in a standard way. Specific "target lesions" are selected, their diameters are summed, and response is categorized according to change over time: complete response, partial response, stable disease, progression. AI can automate the measurement and summation of target lesions; But RECIST's rules (which lesion to choose, how to measure, when to call progression) require rigor, and an automatic "progression" label without verification from the radiologist is a harsh decision that will lead to a change in treatment.
step
AI contribution
trap
verification
Previous study matching
automatic registration
misalignment
Visual confirmation of match
Lesion matching
Finding the same lesion
Mapping different lesion
Anatomical position control
change measurement
Diameter/volume/percentage
Method/protocol difference
Compare with the same method
RECIST total
Target lesion total
Wrong target selection
Compliance with RECIST rule
Response category
draft proposal
Automatic "progression"
The radiologist categorizes
Caution: An automatic "growth" or "progression" output is always verified before changing therapy. The most common cause of false progression is that the previous and current measurement were made with different methods/planes/protocols. An incorrect progression label can unnecessarily change a treatment that is working.
three mini cases
Case 1 — Wrong lesion match. There are two metastases in one liver. Automatic tracking matches the large lesion in the new exam with the small lesion in the previous exam (confusing the anatomical location) and reports "80% growth". The radiologist checks the match: two separate lesions have actually been confused; Both are stable when each is paired with its counterpart. The mismatch produced a spurious progression; The radiologist corrected it.
Case 2 — Different protocol trap. The previous examination of a lung nodule was taken with a 5 mm section, the new one with a 1 mm section. Thin section shows the nodule more clearly and increases the automatic measurement from 6 to 8 mm. The system says "growth". The radiologist notices the protocol difference, re-evaluates with the same slice thickness, and finds that the nodule is stable. The protocol difference had created false growth.
Case 3 — RECIST was implemented correctly. In an oncology patient, the AI automatically calculates the diameter sum of the RECIST target lesions and produces a “partial response” plot. The radiologist checks that the target lesion selection complies with RECIST rules, that the measurements are made with the same method and that there are no new lesions, and confirms the category. The oncologist continues the treatment safely. AI has accelerated the calculation; The radiologist confirmed the clinical decision.
Weak prompt / Strong prompt
Weak prompt:
Decide if the nodule has grown, previous 5 now 8.
No method, plane, protocol, match verification; the model blindly says "grown up" and confirms the false progression.
Powerful prompt:
Your role: ASSISTANT in tracking comparison. Decision making; Create a verification checklist. I'll give you the measurement and context from two studies. Ask and mark the following: (1) is the lesion matching anatomically correct, (2) is the measurement the same method/plane, (3) is the protocol/section thickness the same, (4) is there a new lesion, (5) is the change above the measurement margin of error. If any are in doubt, issue a "consistent remeasurement required for change decision" warning. The decision is up to the radiologist. Previous: 5 mm, axial, manual, 5 mm section. New: 8 mm, axial, automatic, 1 mm section.
Powerful prompt queries match, method, and protocol consistency; The decision remains with the radiologist.
Copiable prompt templates
TRACKING VERIFICATION TEMPLATEYour role: ASSISTANT in tracking comparison. I will give you two examination measurements and context. Check for accuracy of lesion matching, method/plane/protocol consistency, presence of new lesions, and whether the change is within the margin of measurement error. Recommend consistent remeasurement of each questionable item. The decision is up to the radiologist. Data: [write]
RECIST CHECK TEMPLATEI will give you the target lesions, their diameters and dates. In terms of RECIST: calculate the appropriateness of target lesion selection, diameter sum, percentage change and possible response category as DRAFT; Also tick the question "Are there any new lesions?" The final answer category is the radiologist's. Data: [write]
MISMATCH AUDIT TEMPLATERemind me of the points I need to check when auditing automatic lesion matching: is the anatomical location the same, is there confusion with a neighboring lesion, was this lesion really measured in the previous examination. Create a short checklist.
FAKE PROGRESSION IDENTIFICATION TEMPLATE I will give you a "growth/progression" alert and protocol information for two examinations. Evaluate whether this change is real or may be due to method/plane/protocol/contrast phase difference and recommend verification that should be done before changing treatment. The decision is up to the radiologist. Data: [write]
Common mistakes
- Accepting automatic matching without checking. Incorrect lesion matching produces spurious growth/shrinkage.
- Comparing different protocols. The slice thickness/contrast phase difference creates false progression.
- Automatically change treatment with "progression". Response category requires radiologist verification.
- Skipping the new lesion check. In RECIST, a new lesion in itself may mean progression.
- Not to assume that the change is within the margin of measurement error. Small differences may not be real change.
Tip: At every follow-up comparison, ask two questions together: “Am I comparing the right lesion?” and “Was it measured under the same conditions?” If neither is yes, the change you are seeing may not be real but a measurement/match illusion.
In summary
Comparative reading and tracing supplement the meaning of an image with previous studies; It is one of the most valuable jobs in radiology. Artificial intelligence saves time by automatically aligning with the previous examination, lesion matching and change measurement; It can calculate the target lesion sum in frameworks such as RECIST. But there are three major pitfalls: incorrect lesion matching, different protocols, and differences in method—all of which can produce spurious growth or progression. The radiologist checks the match, verifies that the measurements were made under the same conditions, checks the new lesion, and assigns the response category himself. An automatic "progression" must be verified before changing treatment; The final evaluation is the radiologist's.
Application task
Construct a follow-up case (for example, an oncology patient, two examinations). Evaluate match, method and protocol consistency with the "Tracking Verification" template. Then add an intentional decoy — one for a different slice thickness, one for a mismatch of lesions — and test whether the “Distinguish Spurious Progression” template catches them. If you are using RECIST, calculate a target lesion total with the "RECIST Control" template and add the new lesion control.
checklist
- [ ] I checked the automatic lesion matching by anatomical location.
- [ ] I confirmed that the measurements were made using the same method and plane.
- [ ] I checked protocol/slice thickness/contrast phase consistency.
- [ ] I also evaluated whether there was a new lesion.
- [ ] I confirmed that the change is above the measurement margin of error.
- [ ] I verified the automatic "progression" output before the treatment change.
- [ ] I gave the response category (RECIST); I did not transfer it to the model.