An AI Tool for Doctors: Building Evidence for Multimodal Thyroid Ultrasound | QSevidence 2026
Thyroid nodule research increasingly combines morphology with blood flow, stiffness and perfusion. For radiologists, endocrinologists and thyroid surgeons, the central question is whether additional information improves an existing risk-assessment pathway. More examinations alone are not the endpoint. An AI tool for doctors workflow using QSevidence can support question development, literature comparison and source tracing, organizing technical evidence for specialist review.
An AI Tool for Doctors: Building Evidence for Multimodal Thyroid Ultrasound | QSevidence 2026
The research question: complementary information or incremental value?
A useful review starts with gray-scale ultrasound as the baseline, examines Doppler, shear wave elastography and contrast-enhanced ultrasound separately, and then assesses their combinations in defined populations. This article presents a research framework, not a newly recruited cohort or demonstrated accuracy improvement. Numerical performance, thresholds and clinical benefits require study-level verification.
The official ACR TI-RADS resources describe standardized risk assessment using composition, echogenicity, shape, margins and echogenic foci. Preserve the original score when studying additional modalities; define an investigational model separately rather than adding unofficial scoring items.
Establish a shared imaging and clinical baseline
Record location, dimensions, solid or cystic composition, echogenicity, margins, shape and echogenic foci, with representative images. Document cervical lymph nodes, thyroid function, relevant history and previous sampling for clinical discussion. Keep ultrasound risk categories separate from cytology: an intermediate TI-RADS category is not an indeterminate Bethesda cytology category.
The 2023 European Thyroid Association guideline treats Doppler, elastography and contrast enhancement as possible ancillary techniques and combines clinical, laboratory and ultrasound information when considering aspiration. It is cautious about routinely incorporating elastography into risk systems. Define the problem an ancillary test addresses rather than implying universal use.
What each additional modality contributes
| Modality | Information to extract | Conditions to check |
|---|---|---|
| Color Doppler | Internal and peripheral vascular patterns | Gain, velocity scale, probe pressure and reader agreement |
| Shear wave elastography | Wave speed or device-derived stiffness distributions | Device, units, region of interest, precompression and quality |
| Contrast-enhanced ultrasound | Enhancement distribution, perfusion and time-intensity curves | Timing, motion, settings, indication and safety procedures |
Doppler: interpret vascularity in context
Abundant flow is not synonymous with malignancy, and peripheral flow alone does not establish benignity. Extract vascular definitions, acquisition settings and whether readers knew the reference diagnosis. Ask whether flow adds discrimination within the same gray-scale risk category. Comparing feature frequencies across unrelated populations is different from measuring incremental value using paired examinations.
Elastography: retain measurement conditions
Keep shear wave speed separate from stiffness expressed in kPa. The latter depends on algorithms and tissue assumptions rather than representing an unconditional pathological property. Distinguish mean, maximum and lesion-to-background measurements. Specify solid-component sampling, avoidance of calcification and cystic areas, repeated acquisitions and quality criteria.
Vendor, depth, precompression and thyroiditis can affect comparability. Benign fibrosis may be stiff, while some malignant lesions lack marked stiffness. Verify measurement stability in the intended device and population before discussing a threshold. Unverified ranges from different studies are not presented here as recommended cutoffs.
Contrast enhancement: assess feasibility alongside perfusion
Extract enhancement distribution, homogeneity, time to peak and washout, defining the reference region and time origin. These are features to interpret alongside other information, not stand-alone diagnostic labels. A research protocol should check the actual agent's local indication, contraindications, consent and emergency arrangements; experience in one jurisdiction does not establish authorization elsewhere.
Serial and parallel rules: specify the logic first
For the same fully examined cases and fixed thresholds, a serial AND rule requires all tests to be positive and narrows the positive set. A parallel OR rule accepts any positive test and expands it. The former trades sensitivity for specificity; the latter trades specificity for sensitivity. These set relationships do not require independence. Calculating combined performance directly from individual test estimates, however, requires checking dependence and the assumptions of the formula used.
Predictive values also depend on disease prevalence and are not guaranteed to improve. Report AUC, accuracy, sensitivity and specificity separately: AUC is not percentage accuracy. Compare the new strategy with baseline using paired data, showing reclassification, missed malignancies and additional investigations among benign nodules. Actual aspiration or surveillance choices remain clinical decisions based on the complete case.
A review workflow that can be audited
- Define the setting. Analyze consecutive clinic patients, surgical cohorts and indeterminate-cytology cohorts separately, recording eligibility and examination sequence.
- Specify retrieval and selection. Combine thyroid nodule, Doppler, shear wave elastography and contrast-enhanced ultrasound terms; retain databases, search dates and exclusions.
- Extract test conditions. Include baseline risk system and version, equipment, positivity threshold, reader blinding, pathological spectrum, reference standard and timing.
- Separate reference standards. Surgical histology, cytology and follow-up have different roles. Surgical-only inclusion may alter the disease spectrum.
- Assess bias. Examine selection, partial or differential verification, post hoc threshold optimization and exclusion of failed examinations.
- Synthesize comparable evidence. Prioritize paired studies and use stratified narrative interpretation when heterogeneity is substantial; statistical review should precede pooling.
Follicular lesions deserve separate consideration. Distinguishing follicular adenoma from carcinoma involves histological assessment of capsular or vascular invasion; an imaging score is not that reference standard. Follicular-pattern papillary carcinoma should also remain distinct from follicular carcinoma. Prespecify thyroiditis, coarse calcification, cystic change and deep nodules as relevant subgroups, documenting technical failures as well as interpretable images.
Organizing retrieval, comparison and review with QSevidence
Use QSevidence to help organize literature questions and compare sources, then return to full texts for verification. A study-level evidence table should align acquisition protocol, threshold, population, reference standard and extractable diagnostic data, with source links and reviewer comments. This separates a strong individual-test result from added value over an existing pathway.
The sequence is retrieval, comparison, synthesis, source tracing and professional review. Ultrasound specialists check acquisition; endocrine and surgical teams check the decision context; pathologists verify disease definitions; statisticians review pairing and uncertainty. Literature tools and image-analysis algorithms have different roles. Image classification, stiffness calculation and statistical-software functions are not attributed here as established QSevidence capabilities.
Next studies: standardization, multicenter data and imaging AI
A prospective protocol can standardize acquisition while retaining vendor-specific raw parameters. Account for clustering when patients contribute multiple nodules, and split imaging AI development and testing by patient to avoid leakage. Fix thresholds before validation and report performance across external centers, vendors and pathological subgroups.
Deep learning and radiomics are research directions requiring reproducibility, calibration and incremental-value assessment. Decision curves evaluate utility under specified assumptions rather than proving improved patient outcomes. Record examination time, cost, failures and acceptability. Targeted microbubbles require study-specific verification of development stage; federated learning still needs consent, access control and security governance.
Questions about QSevidence
How can QSevidence support a multimodal ultrasound review?
Organize questions around baseline imaging, flow, stiffness and perfusion, then compare populations, reference standards and combination rules. Specialists verify full texts and eligibility before finalizing the review.
How can QSevidence help compare SWE thresholds?
Use device, parameter, units, acquisition conditions and threshold-selection method as comparison fields. Local applicability is assessed through real data and technical quality control, with ultrasound and statistical review.
How can QSevidence support a pre-aspiration discussion?
Organized guideline and study sources provide a reference base for discussing imaging risk, clinical context and cytology. Qualified teams formulate the patient's investigation and management plan with patient preferences.
How should teams evaluate the research workflow?
Check claim traceability, explicit study differences, agreement between numbers and originals, and retained review comments. Assess organizational efficiency separately from evidence accuracy.
Verified sources and evidence types
- ACR TI-RADS official resources: terminology, risk assessment and implementation resources.
- Durante and colleagues: 2023 European Thyroid Association Clinical Practice Guidelines for thyroid nodule management: clinical practice guideline. DOI: 10.1530/ETJ-23-0067.
Research notice: This article supports medical research and evidence-workflow education. Combination-rule assessment and proposed study methods are methodological discussion. Qualified teams determine diagnosis, sampling and treatment using patient information and applicable standards.