When scans and blood tests disagree, AI could help oncologists decide what to do next

When a cancer patient’s scans and a blood test for tumor DNA tell different stories, doctors have long faced an uncomfortable question: which one is right? A new review argues that this is the wrong question altogether. Writing in Frontiers in Oncology, researchers from Zhejiang Provincial Hospital of Chinese Medicine propose that disagreement between imaging and circulating tumor DNA (ctDNA) testing is not an error state to be resolved in favor of one modality, but a structured form of uncertainty that can be classified, quantified, and ultimately managed with the help of artificial intelligence.

The review, a focused narrative synthesis rather than a systematic review or meta-analysis, draws on a structured PubMed/MEDLINE search updated through 17 September 2026. Its central premise is that imaging and ctDNA measure fundamentally different dimensions of cancer. Imaging is spatially resolved and lesion centered: it shows where disease sits, how much of it there is, and how it changes shape over time. ctDNA, by contrast, is a systemic molecular signal weighted toward lesions that shed DNA into the bloodstream, and it can reveal molecular evolution, such as emerging resistance mutations, that no scan can see. Because these are non-identical measurements, the authors argue, some disagreement is biologically expected rather than anomalous.

That expectation is borne out by quantitative evidence from several disease settings. In a cohort of 821 patients with nasopharyngeal carcinoma, MRI-based response and post-treatment cell-free Epstein-Barr virus DNA response were discordant in 25.3 percent of patients. In a pilot cohort of advanced genitourinary malignancies treated with immune checkpoint inhibitors, ctDNA dynamics and imaging agreed in 15 of 18 evaluable patients, but among the three discordant cases, two later developed central nervous system metastases and one developed extracranial progression despite persistently undetectable ctDNA. In postoperative pancreatic cancer surveillance, imaging and ctDNA were concordant in 39 of 50 patients, and 10 of the 11 discordant cases had radiologic recurrence with negative ctDNA. These examples also illustrate why the review cautions against quoting any universal discordance rate: tumor type, disease burden, metastatic distribution, assay design, and sampling intervals differ substantially across studies.

To bring order to this heterogeneity, the authors propose a multidimensional, multilabel taxonomy of discordance spanning five dimensions: directional, temporal, spatial, molecular, and technical. The dimensions are deliberately not mutually exclusive. A single patient might simultaneously show directional discordance, with imaging and ctDNA moving in opposite directions; temporal discordance, with a ctDNA rise preceding radiographic progression; molecular discordance, with an emerging resistance clone invisible on scans; and spatial discordance, with a lesion whose behavior is poorly represented in plasma. A technical dimension captures mismatches plausibly explained by analytical or preanalytical effects, such as low-level ctDNA oscillation near the assay’s limit of detection or borderline lesion change near a RECIST threshold. The taxonomy is explicitly author-proposed and intended for prospective testing, not a validated clinical instrument.

The review also insists on operational definitions that researchers can apply consistently. Isolated discordance would denote a single discrepant observation not yet reproduced, while persistent discordance would require confirmation on at least two sequential assessments using comparable methods. The authors stress that these are proposed definitions, not validated thresholds, and that no universal acceptable interval between imaging and blood collection can currently be justified across tumor types, treatment mechanisms, and assay platforms. They likewise reject any universal percentage change in ctDNA as a definition of molecular response or progression, arguing that interpretation must be assay-specific and account for analytical variation, baseline abundance, and the biological meaning of the tracked alteration.

Perhaps the review’s most consequential argument is the separation of three evidentiary steps that are often conflated: molecular detectability, prognostic validity, and treatment actionability. Prognostic complementarity is well supported. In metastatic non-small-cell lung cancer, a longitudinal ctDNA model derived from the IMpower150 trial stratified overall survival even within RECIST response categories: among patients with stable disease, median overall survival was approximately 7.1 versus 22.3 months, and among those with partial response it was approximately 8.8 versus 28.6 months. Yet, as the authors emphasize, such findings show that molecular trajectories add information within radiographic categories; they do not show that acting on them improves outcomes.

Randomized evidence cuts both ways on actionability. The PADA-1 trial provides randomized support that molecular-first treatment adaptation can be beneficial in a narrowly defined setting: patients with advanced hormone-receptor-positive, HER2-negative breast cancer receiving an aromatase inhibitor plus palbociclib who developed a rising ESR1 mutation without synchronous progression benefited from a biomarker-guided switch of endocrine therapy. The DYNAMIC trial showed that postoperative ctDNA can guide adjuvant treatment allocation in stage II colon cancer while reducing chemotherapy use without compromising recurrence-free survival. But ALTAIR, embedded within the CIRCULATE-Japan program, randomized 243 patients with resected colorectal cancer who became ctDNA positive after standard therapy yet had no radiological disease; trifluridine/tipiracil did not significantly improve disease-free survival versus placebo, with a median of 9.30 versus 5.55 months and a hazard ratio of 0.79. Molecular recurrence, in other words, does not automatically make an early intervention beneficial.

Against this backdrop, the authors frame multimodal AI not as an automated treatment selector but as an uncertainty-aware decision-support tool. They propose requirements for future systems rather than describing a validated architecture: such systems should preserve modality-specific information rather than fusing signals into a single label, model irregular longitudinal sampling using exact timestamps, classify discordance with multilabel outputs, quantify uncertainty and permit abstention when data quality is insufficient, generate ranked biological hypotheses rather than assert causality, and identify the next assessment most likely to reduce clinically relevant uncertainty. Candidate methods include joint longitudinal and time-to-event models, architectures designed for irregularly sampled time series, calibrated probabilistic and conformal prediction, and value-of-information approaches, though the authors note that comparative evidence is insufficient to prescribe any single architecture.

To translate classification into clinical behavior, the review proposes four clinical action tiers, distinct from its five-level evidence maturity framework. Tier A applies to isolated, low-confidence discordance and supports protocol-based follow-up. Tier B applies to reproducible or concerning discordance that remains insufficient for intervention and supports repeat ctDNA testing, a shortened imaging interval, or technical review. Tier C applies to persistent, higher-confidence discordance and could justify lesion-directed imaging, multidisciplinary review, or biopsy. Tier D, evidence-supported action, applies only when a validated disease-specific biomarker-intervention strategy supports treatment modification. In a worked vignette, a patient with stable disease on CT whose ctDNA rises reproducibly on two sequential assessments with an emerging resistance alteration would fall within Tiers B and C, warranting reassessment and confirmation but not automatic systemic treatment switching.

The authors are candid about what remains unproven. Level 5 of their evidence maturity framework, prospective clinical utility, has been achieved only in selected biomarker-specific settings and has not been established for any general discordance-aware imaging-ctDNA AI system. They catalog potential harms that future evaluation must weigh, including false-positive molecular progression, unnecessary imaging or invasive procedures, anxiety, financial toxicity, premature discontinuation of effective therapy, and false reassurance from undetectable ctDNA, alongside unequal model performance across institutions, assay platforms, and patient populations. They call for prospective validation that prioritizes net benefit, harms, transportability, governance, and cost rather than discrimination alone, and for randomized trials of treatment changes only when a specific intervention is biologically plausible and supported by sufficient preliminary evidence. The field, they conclude, should advance only when prospective studies demonstrate that discordance-aware decision support improves patient-relevant outcomes without creating excess procedures, treatment switching, cost, or inequity.

Subject of Research: Integrating longitudinal imaging and ctDNA monitoring through a proposed discordance taxonomy and uncertainty-aware multimodal AI in precision oncology

Article Title: From discordance to clinical decision support: integrating longitudinal imaging and ctDNA through multimodal artificial intelligence in precision oncology

Article References: Zheng, J., Chi, L., & Guo, Y. (2026). From discordance to clinical decision support: integrating longitudinal imaging and ctDNA through multimodal artificial intelligence in precision oncology. Frontiers in Oncology, 16, Article 1963645. https://doi.org/10.3389/fonc.2026.1963645

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1963645

Keywords: circulating tumor DNA, liquid biopsy, imaging, discordance, multimodal artificial intelligence, precision oncology, treatment response, clinical decision support, molecular residual disease, biomarker-guided therapy, RECIST, prospective validation