Abstract: Pairing of non-contrast chest CT with CTPA/SPECT V/Q at near-time points in the same patient can be used to study whether the model can identify clues from limited images; the pairing design must account for changes in time and clinical status.
*Figure 1: The evaluation closed loop organized around the requirements of this attachment is a method design illustration, not a project result.*
The medical and life sciences team is shifting from "is the model capable" to "is the model reliable within what tasks, what populations, what data conditions, and what risk boundaries?"Pairing of non-contrast chest CT with CTPA/SPECT V/Q at the same patient at a near-time point can study whether the model can identify clues from limited images; the pairing design must account for changes in time and clinical status. For LumeSage, professional data, expert production and continuous evaluation need to form the same closed business loop: the problem is clearly defined, the data and reference standards are traceable, the expert contribution can be verified, the model operation can be reproduced, and the results can be returned to delivery and the next round of improvement.
This article extracts demand signals based on the attachment "pe-paired-imaging-reference", and combines public research, medical AI reporting specifications and engineering practices as of October 2026 to propose an evaluation design suitable for scientific research and project delivery. The numbers, fields, and acceptance rules in the attachment are all requirement specifications or examples, and do not mean that the data has been purchased, annotated, met compliance requirements, or that the algorithm is effective.
1. Define the business problem first, rather than buying data first
When the business party asks for a data set, what is actually purchased is usually not the file, but the ability to answer a decision-making question. For this topic, we must first clearly describe who will use the model, what the input is, which step of work the output will affect, what consequences the error will bring, and where manual review will occur. The appendix corresponds to adult acute pulmonary embolism; the sample requirement for the phased QC in the appendix is 5–10 cases, which is not the overall cohort size.
recommends splitting the research objectives into tasks that are irreplaceable for each other: DICOM geometry/layer thickness/coverage, reporting, reference checking, time difference and treatment intervention information. Each task separately stipulates the target user group, applicable scenarios, excluded scenarios, allowed inputs, expected outputs, rejection conditions and risk levels. In this way, the business team can judge whether the sample covers real decisions, the algorithm team can choose classification, retrieval, segmentation, sorting or prediction tasks, and medical experts can also judge whether the reference standard is established.
This split directly affects commercial viability. During the project establishment phase, a small-scale feasibility assessment of available data, authorization paths, expert supply, quality costs, and verification objectives can be conducted first, and then the batch size and stage threshold can be determined. Small-scale QC only verifies data applicability and is not enough to prove model performance or clinical effectiveness; therefore, external communication should list “research data availability”, “technical indicator compliance” and “effective clinical workflow” as different milestones.
2. Translate attachment specifications into computable data contracts
The data contract should itemize entities, fields, types, units, encoding versions, allowed values, null meanings, source systems, timestamp definitions, association keys, and acceptance rules. For non-enhanced CT and CTPA within 7 days or lung SPECT V/Q, the same patient image and report must not only check whether the file can be opened, but also save the spatial geometry, acquisition conditions, source documents and conversion process; for DICOM geometry/slice thickness/coverage, reports, reference checks, time differences and treatment intervention information, it must also indicate whether the label comes from the original record, expert interpretation or algorithm derivation. The specific requirements in the
attachment can be converted into machine-checkable constraints: paired index → time window review → clinical change marking → blind interpretation → external multi-center validation. Hard conditions and recommended conditions should be separated, and missing values should be distinguished between "not applicable", "not collected", "unavailable" and "parsing failure", otherwise the downstream model will combine different reasons into the same value. It is best for rules to have example lines, counterexamples, and automatic verification results, and keep a version of the rule with each batch.
A valid data contract also needs to define sampling, tolerance, rejection, repair and re-inspection. For example, patient-level association is not just about checking whether the ID is non-empty, but also checking whether imaging examinations, reports, pathology, and follow-up refer to the same anonymous object; for the complete matching rate, image technology qualification rate, reference standard determination rate, and patient-level hierarchical diagnostic efficiency, the denominator and statistical caliber must be clear. The purpose of acceptance is not to count every anomaly as a failure, but to identify whether the anomaly will undermine the target mission.
3. How to establish reference standards, labels and expert judgment
There is rarely an absolute “gold standard” for medical evaluation that is divorced from the scene. Reference standards must state the source, date, edition, expert qualifications, independence and dispute resolution. For the DICOM geometry/layer thickness/coverage, reports, reference examinations, time differences and treatment intervention information of this topic, concurrent examinations, pathological results, formal diagnoses, scales and expert consensus can be stored in layers, and the chain of evidence should be retained; multiple evidences should not be forced into one label and then discard inconsistent information.
The labeling process should be divided into rule learning, independent labeling, disagreement review and arbitration. Dual independent annotation is not to create a majority vote, but to measure the decidability of the task itself. For difficult, low-quality or insufficient data samples, it should be allowed to mark "unable to determine", "insufficient evidence" and "out of range", and the reasons should be recorded. This avoids misinterpretation of clinical uncertainties as supply-side deficiencies or model errors.
Evaluation questions or cases should also form a data card: source type, inclusion and exclusion, demographic and disease composition, time range, missing modalities, equipment differences, anonymization method, permission scope, label definition, sampling results and limitations. Use these fields to connect paired index → time window review → clinical change mark → blind interpretation → external multi-center verification, so that third parties have the opportunity to review the applicable scope of the benchmark instead of just seeing an accuracy ranking list.
*Figure 2: Schematic of the quality dimensions and decision-making nodes of this topic.*
4. Evaluation set and operating protocol: Make scores reproducible The
protocol freezes the test object and scoring logic before running the model. At least record the model name and version, weight or API version, system prompts, temperature, random seed, tool permissions, input truncation method, retry strategy, time and running environment. For non-enhanced CT and CTPA within 7 days or lung SPECT V/Q, the same patient imaging and reporting tasks must also be fixed with image preprocessing, coordinate/direction convention, resampling method, available time window and missing modality strategy.
Data segmentation should be based on patients as the smallest unit; when there are multiple centers, out-of-center tests must be prepared, and out-of-time tests must be set up when there are time evolution issues. Different slices, follow-up or repeated videos of the same patient cannot be randomly divided into training and test sets. If the attachment design is a positive cohort, the test results can only be interpreted as the discrimination ability of this cohort and a given task, and the positive predictive value or screening value in the population cannot be directly deduced.
The blind test set should be kept by an independent role to avoid training, parameter adjustment, prompt word engineering or manual rules from touching the test answers. It is recommended to report the public development set, internal validation set, private set-aside set, external validation set and prospective validation separately. Run lists and hashes are generated for each list update to reproduce matching completeness rates, image technical qualification rates, reference standard adjudication rates, patient-level stratified diagnostic performance, and whether tools, manual assistance, or multiple sampling are used for the model. The
5. Quality indicators should cover effectiveness, stability and safety The
indicator should start from the target loss and should not first pick a score that is easy to get high. The classification task can report sensitivity, specificity, AUROC, AUPRC, calibration and decision threshold; the segmentation task should report Dice, surface distance, connectivity, volume deviation and small structure performance; the generative answer needs to examine factual correctness, evidence support, omissions, harm, refusal and expert scoring. Corresponding to this topic, priority should be presented: matching complete rate, image technology qualification rate, reference standard determination rate, and patient-level hierarchical diagnostic efficiency.
All core indicators should be given uncertainty intervals and report sample size, category distribution, statistical units and confidence interval algorithm. Case-level, lesion-level, frame-level or voxel-level numbers cannot be mixed together. Macro-averaging and micro-averaging answer different questions. Positive enrichment will change the accuracy, and threshold changes will change the sensitivity and false positives. Therefore, the list can display the main score, but also provide safety sub-scores, task sub-scores, group sub-scores, resource consumption and failure types.
"Expert consensus" and "model correctness" cannot be regarded as the same concept. Inter-expert agreement is a reference for measuring task decidability; when a model disagrees with a single expert, it is necessary to go back to the evidence and analyze whether it is expert differences, insufficient label definition, or model errors. For DICOM geometry/layer thickness/coverage, reporting, reference checks, time differences and treatment intervention information, a dual evaluation can be used: one part is an objectively verifiable indicator, and the other part is evaluated by blinded experts and published scoring anchors and consistency data.
6. Bias, external generalization and failure analysis The
model may use cues outside of the target to achieve a superficially high score. For example, changes in condition/treatment during the interval, non-equivalence of different examinations, case selection bias, and differences in scanning protocols. Error analysis therefore examines the site, device, how the task is performed, instrument templates, language, age group, gender, disease severity, lesion size, and missing modalities; any features that are identifiable or should not be used should not serve as shortcuts.
Generalization assessment requires progressively increasing levels of difficulty: same-center patient-level set-aside, out-of-time validation, device/hospital external testing, different workflows, different populations, and missing modality testing. Center-level confidence intervals or distributions should be given for results across centers, not just aggregated scores. A fair audit should first define the groups relevant to the expected deployment, and then check whether the sample is sufficient to support reliable conclusions; when the sample is insufficient, it should be marked as "insufficient evidence" and cannot be covered up by the overall average.
Error analysis should include actionable failure classification, such as changes in condition/treatment between intervals, non-equivalence of different examinations, case selection bias, and differences in scanning protocols. Each failure should be linked to input quality, label source, preprocessing, model output, and expert review. High-risk errors, indecisions, and model rejections need to be presented separately. In particular, for mental and psychological scenes, it is necessary to avoid directly assigning facial behaviors to disease meanings; for tumor prognosis, time leakage should be prevented; for emergency imaging, the relative risks of missed detection and false positives should be explained.
7. From small sample pilot to delivery acceptance
It is recommended to set up small batch verification and phased acceptance first, rather than arguing about the full amount of data at the end of the project. The first stage checks the source, license, field mapping, file format and annotation manual; the second stage draws representative samples to check the completeness and annotation consistency; the third stage runs the target model or baseline to confirm whether the data supports the proposed task. The sizes and specifications provided in the attachment are requirements input and do not represent delivered results.
Delivery package may contain data dictionary, patient/examination index, original or authorized derived files, labels, data cards, processing logs, calibration reports, sampling records, error lists, change records, and retest results. For paired index → time window review → clinical change marking → blind interpretation → external multi-center validation, it should be clear which items are blocking defects, which ones can be reworked proportionally, and which ones need to be adjudicated by medical experts. Each rejection should be located to an anonymous sample, field, or file, with rules and evidence attached. The
The business contract should also stipulate usage boundaries, data retention period, access control, de-identification responsibilities, problem response time, rework batches and acceptance signature roles. If the true source or ethical conditions are not met, no matter how good the algorithm is, it should not enter the follow-up process. It is best to issue acceptance results separately by case, modality, field, label, document and legal authorization to avoid "overall qualified" covering up key defects.
8. Continuous evaluation and business closed loop
benchmark is not a one-time list, but a versionable measurement system. Confirm tasks, data and model snapshots before running; save input, output, tool calls, failure status and running logs during running; and review them by independent scorers and experts after running. According to the paired index → time window review → clinical change marking → blind interpretation → external multi-center verification precipitation question list, the failure mode can be converted into the basis for the next round of data supplement, label revision or model improvement.
Continuous evaluation should distinguish between baseline changes and model changes. When adding new cases, modifying labels, adjusting scoring weights, or replacing model interfaces, a version number and change description should be generated, and bridging tests should be conducted on the old and new versions. Otherwise, the increase or decrease in the ranking score may just be due to the change of test set. Public reports display vetted task statistics, and sensitive test answers are placed in a permission-controlled benchmarking service using access logs and anti-duplicate query mechanisms.
The business closed loop can be promoted around "demand clarification-sample feasibility-expert production-data quality control-private evaluation-error attribution-improvement retest-batch acceptance". For pairing completeness rate, image technical qualification rate, reference standard decidability rate, and patient-level hierarchical diagnostic performance, map each failure category to data supplementation, annotation training, model engineering, or usage restrictions to measure whether the improvement holds on an independent test set. LumeSage can provide data engineering, expert annotation, quality control and Medical Gate private evaluation work packages based on authorization and contract. The specific team, scale and time need to be confirmed on a project-by-project basis.
9. Work packages and capability boundaries that LumeSage can undertake
LumeSage's business positioning should be in the auditable service chain, rather than exaggerating the "algorithm effect": develop a data plan around non-enhanced CT and CTPA within 7 days or lung SPECT V/Q, the same patient image and report, establish a field dictionary and data card, coordinate expert annotation that meets the authorization requirements, perform format/association/content quality control, and then use the model and test protocol agreed upon by the project to conduct Medical Gate private evaluation. Each step should leave evidence of input and output, responsible roles, and versions. Before starting the
project, LumeSage needs to work with the customer to confirm data rights, hospital authorization, ethical review or informed consent applicability, desensitization plan, expert qualifications, delivery scope, set-aside management and acceptance indicators. Adult acute pulmonary embolism in the attachment; the phased QC sample requirement in the attachment is 5-10 cases. The non-overall cohort size is only used as a reference for this plan discussion. It cannot be used to claim to the market that LumeSage has available data of this scale, nor can the research service be described as medical device certification or clinical diagnostic services.
Deliverables are not just sample files, but can also include usability assessments, data dictionaries, task definitions, annotation specifications, spot inspection reports, error stratification, evaluation protocols, and versioned results reports. Its value is to reduce the friction of the algorithm team in early clarification, repeated cleaning and non-reproducible evaluation, and to help the business side determine whether the next investment should be added to collection, labeling, model or verification. The actual scope of services shall be subject to the written work description and contract.
10. Research and industrial development prospects
Future technical directions may include temporal basic models, multi-modal representation learning, robust fusion of missing modalities, retrieval enhancement, active learning, expert feedback and uncertainty calibration. But architectural frontiers do not automatically lead to clinical validity; models must be compared against transparent baselines, manual processes, and appropriate reference standards on predefined tasks. For non-enhanced CT and CTPA within 7 days or lung SPECT V/Q, the same patient images and reports should especially be evaluated for modality loss, acquisition differences, and input quality changes.
Data construction can be upgraded from static labels to event and evidence maps: recording who made judgments when and based on what data, how the data was transformed, what time cutoff the model saw, and how the output was accepted or corrected by experts. Only by retaining the sources and differences of DICOM geometry/layer thickness/coverage, reports, reference checks, time differences and treatment intervention information can subsequent algorithm research distinguish biological signals, process characteristics and measurement errors. Synthetic data may be used for development and stress testing but cannot be used as a substitute for real patient data without independent validation.
The maturity of industry evaluation is ultimately reflected in the transparent scope of application, clear recurrence process, traceable risks, explainable failures, and auditable updates. It is recommended that priority should be given to establishing benchmarks with narrow tasks, strong evidence, and clear reference standards, and then gradually expand to cross-disease types and cross-institutions. For tasks with insufficient evidence, a reasonable conclusion should be to supplement data or design a study rather than force a ranking.
Method Boundaries and Data Governance
This article implements "method boundaries and data governance" into the actual data process of non-enhanced CT and CTPA within 7 days or lung SPECT V/Q, the same patient imaging and reporting. The core is not to add more terms to the document, but to allow each rule to correspond to a data check, an expert judgment, or a reproducible model measurement. For acute pulmonary embolism in adults; the appendix phase QC sample requirement is 5–10 cases, which is not the overall cohort size. The included population and uncovered scenarios must also be stated. The
team can refine the pairing index → time window review → clinical change marking → blind interpretation → external multi-center verification into roles and products: data engineers maintain the structure and traceability, clinical experts define labels and ruling boundaries, algorithm personnel freeze tasks and run configurations, quality control personnel verify sampling and deviations, and project leaders sign acceptance. If these roles are performed by the same person, this should be disclosed in the report and an independent review should be added. When forming the conclusion,
simultaneously reports the matching complete rate, image technology qualification rate, reference standard determinability rate, patient-level stratified diagnostic efficiency, limitations "Small-scale QC only verifies the applicability of the data, which is not enough to prove model performance or clinical effectiveness", and points out that "changes in condition/treatment during the interval, non-equivalence of different examinations, case selection bias, and differences in scanning protocols" need to be solved by subsequent research. This allows the reader to understand what the benchmark can answer and what it cannot infer. Public articles only describe the methodological framework and do not publish client data, identifiable cases, or unauthorized original material.
Conclusion: Turn data delivery into verifiable capabilities
This attachment reminds the industry: Acute Pulmonary Embolism Paired Image Evaluation: How to establish a reliable reference between non-enhanced CT and CTPA. High-quality data for medical AI does not end with the number of files or a single score, but is based on task matching, credible sources, traceable expert judgment, reproducible model evaluation, and clear risk boundaries. LumeSage can combine solution evaluation, data processing, expert annotation, quality control and Medical Gate private evaluation into phased projects based on customer authorization, data cards and contracts; whether the project is feasible, whether the data is available, and how many experts and time are required must be verified in the specific scenario.
Method Boundary and Privacy Description
This article is a research and data engineering method analysis and does not constitute diagnosis, treatment recommendations, medical device performance certification, or legal advice. Psychiatric videos and images are highly sensitive data; case-level information must not be disseminated publicly. Any training, validation or benchmarking use should be subject to lawful authorization, applicable ethical review, minimum necessary processing, de-identification, access controls and use limitations. Regarding the paired image evaluation of acute pulmonary embolism proposed in the attachment: how to establish a reliable reference between non-enhanced CT and CTPA, the source, authorization, technical feasibility and acceptance criteria still need to be confirmed by all parties in the project before implementation.
Policy timelines and standard status must be kept distinct. A 2025 implementation opinion from China’s National Health Commission and partner agencies set a 2027 objective for establishing health-sector high-quality datasets and trusted data spaces; the National Data Administration’s 2026 sector action plan sets an end-2028 objective for application-validated datasets. The publicly available Quality Evaluation Specification for High-Quality Datasets is still marked GB/T XXXXX—XXXX (draft for comments). Its three 90-point thresholds are draft provisions, not an effective universal mandate. Check the finalized standard and the specific call before applying.
References and Methodological Basis
[1] RSNA Pulmonary Embolism CT Dataset. Radiology: Artificial Intelligence, 2021. https://doi.org/10.1148/ryai.2021200254
[2] Digital Contrast CT Pulmonary Angiography Synthesis from Non-contrast CT for Pulmonary Vascular Disease. arXiv, 2025. https://arxiv.org/abs/2510.21140
[3] External validation of a commercial AI system for PE detection on chest CTPA: a multicenter study. 2026. https://pmc.ncbi.nlm.nih.gov/articles/PMC13012949/
5–10 samples are only used for preliminary QC
The pulmonary embolism attachment lists non-enhanced chest CT and CTPA or lung SPECT V/Q within 7 days of the same patient as paired examinations, and requires that the two scans correspond to radiology reports.The phased request is for 5-10 samples, which are used for preliminary quality assessment; this number can only test whether the data can be read, whether it is paired, whether the report is complete, and whether it meets the technical conditions, and cannot be used to verify the size of the algorithm training set or diagnostic performance.If treatment, condition changes, or new events occur within the interval between reference examinations, the timeline and clinical background should be retained and potential label instability flagged during analysis, rather than treating the two examinations as the same time point.
