Metastasis AI Agent: Multi-Scale Detection

• Metastasis AI Agent is a family of AI systems that detect, characterize, and predict metastatic disease by integrating multi-scale data from molecular to systemic levels.
• These systems utilize iterative lesion detection, agent token aggregation, and human-in-the-loop workflows to enhance diagnostic precision.
• They demonstrate strong potential in staging, prognostication, and response assessment, though challenges remain in standardization and clinical validation.

“Metastasis AI Agent” denotes a family of AI systems that detect, characterize, predict, or reason about metastatic disease, often while coordinating multiple computational modules or interacting directly with clinicians. In its most explicit formulation, the term refers to “a multi-scale metastasis scoring system that characterizes tumor progression across the initiation, dissemination, and colonization phases by integrating molecular, cellular, and systemic signals” (Wang et al., 27 Aug 2025). In the broader literature, closely related systems include interactive liver-metastasis assessment frameworks, whole-slide pathology detectors, web-based metastasis prediction agents, intra-operative guidance systems, and evidence-grounded theranostic agents for metastatic disease (Hunt et al., 2019, Li et al., 2018, Xu et al., 2022, Schnelldorfer et al., 2023, Chen et al., 14 Mar 2026).

1. Conceptual scope and representative formulations

The term is used in more than one sense. In the Full-Body AI Agent framework, the Metastasis AI Agent is a specialized disease-focused instantiation that coordinates seven biological-level agents—Molecule, Organelle, Cell, Tissue, Organ, Organ System, and Body System—to generate phase-specific metastasis scores (Wang et al., 27 Aug 2025). In clinical imaging, it appears as an interactive assistant that detects lesions, explains predictions, supports measurement, and structures reports for clinician review rather than acting as a fully automated black box (Hunt et al., 2019). In pathology, it can denote either a complete workflow system for metastasis detection and staging or an internal architectural mechanism such as trainable “agent tokens” used to aggregate weakly supervised whole-slide features (Ling et al., 2024). In clinical informatics, it can also denote a conversational web application that predicts long-horizon breast cancer metastasis risk and trains new models on user-provided data (Xu et al., 2022).

Representative formulation	Representative paper	Core function
Multi-scale metastasis scoring system	(Wang et al., 27 Aug 2025)	Scores initiation, dissemination, colonization
Interactive imaging assistant	(Hunt et al., 2019)	Detects, explains, measures, reports liver metastases
Pathology detection and staging system	(Lee et al., 2018)	Detects lymph-node metastases and predicts pN-stage
WSI agent aggregator	(Ling et al., 2024)	Uses trainable agent tokens and mask denoise
Web-based prognostic agent	(Xu et al., 2022)	Predicts 5-, 10-, or 15-year breast metastasis
Evidence-grounded theranostic agent	(Chen et al., 14 Mar 2026)	Predicts RLT response in metastatic prostate cancer

A common misconception is that an agent in this area must be a fully autonomous diagnostic system. The literature shows a broader reality: some metastasis AI agents are explicitly human-in-the-loop, some are workflow orchestrators, and some are internal representation-level modules embedded inside a larger model (Hunt et al., 2019, Ling et al., 2024).

2. Architectural patterns and operational loops

A recurring pattern is the move from one-shot prediction to iterative, structured decision support. In the liver-metastasis framework, the system comprises a lesion detection module, an interpretability module based on Shapley values and HyperSHAP, a natural language generation module, a clinician interaction interface, and an automated measuring system. The workflow is explicit: CT images are ingested, lesion candidates are detected, explanations are generated, an interactive report is produced, the clinician confirms or rejects detections and may add missed lesions, and the validated lesion set is then measured for count, location, and diameter (Hunt et al., 2019). This makes the agent’s role proposal–justification–revision rather than silent automation.

In whole-slide pathology, the same agentic logic appears with different computational primitives. NCRF combines a CNN feature extractor with a fully connected CRF over a $3 \times 3$ grid of patch embeddings, then performs mean-field inference implemented as recurrent layers so that patch predictions become spatially coherent before whole-slide probability maps are assembled (Li et al., 2018). A separate breast-cancer pipeline extends patch-level metastasis detection to patient-level pN-stage by chaining ROI extraction, a ResNet101 patch detector, overlap-tiled heatmap generation, feature extraction from candidate regions, a random-forest lymph-node classifier, and rule-based aggregation across five lymph-node slides (Lee et al., 2018). In weakly supervised WSI analysis, AMD-MIL replaces full self-attention with trainable agent tokens that aggregate information from all instances and then broadcast refined information back through a mask denoise mechanism, yielding weighted bag features for slide classification and ROI localization (Ling et al., 2024).

Agentic orchestration becomes even more explicit in theranostic settings. TheraAgent uses a Radiologist Agent, Biochemist Agent, and Oncologist Agent to extract structured features from PET/CT reports, laboratory tests, and clinical text; an Integrator Agent performs confidence-weighted consensus; SEA-Mem stores prior cases as outcome-annotated memory; an Evidence Agent retrieves VISION/TheraP trial evidence; and a Reasoning Agent fuses profile, memory, and trial evidence into a final prediction (Chen et al., 14 Mar 2026). This suggests that “agent” in metastasis research often denotes a control architecture that coordinates specialized submodels, data sources, and reasoning steps rather than a single classifier.

3. Modalities, targets, and representation spaces

Metastasis AI agents span a wide range of input modalities and prediction targets. In radiology, one line of work uses contrast-enhanced CT to segment liver metastases in colorectal liver metastasis patients, with 52 patients and 1380 liver metastases annotated at voxel level for supervised training (Hunt et al., 2019). In longitudinal neuro-oncology, an automated pipeline curated 896 brain metastases in 177 patients monitored for more than 360 days at approximately two-month intervals after stereotactic radiosurgery, enabling trajectory clustering and 12-month lesion-level response prediction from pre-treatment and first follow-up MRI (Kuhn et al., 8 Sep 2025). In intra-operative decision support, routine laparoscopy images of biopsied and non-biopsied peritoneal lesions are used for lesion detection and benign-versus-metastatic classification during gastrointestinal cancer staging (Schnelldorfer et al., 2023).

In computational pathology, the dominant representation is the WSI as a bag of patches. NCRF uses $768 \times 768$ regions subdivided into $3 \times 3$ grids of $256 \times 256$ patches on Camelyon16, with ResNet-18 or ResNet-34 embeddings linked by a CRF (Li et al., 2018). The pN-stage system uses 256 × 256 patches at 0.243 $\mu$m/pixel with 50% overlap, then aggregates heatmap-derived features to classify each lymph-node slide into Negative, ITC, Micro, or Macro before inferring patient-level pN-stage (Lee et al., 2018). The cleaned Camelyon$^+$ benchmark reprocessed 1,399 WSIs from Camelyon-16 and Camelyon-17, removed low-quality slides, corrected labels, added expert pixel annotations for previously unreleased test slides, and produced a four-class dataset of 1,350 WSIs with 871 negative, 174 micro-metastasis, 251 macro-metastasis, and 54 ITC slides (Ling et al., 2024).

Other metastasis AI agents operate over structured clinical or molecular spaces rather than images alone. iMedBot uses categorical clinical predictors from LSM datasets to predict 5-, 10-, or 15-year breast cancer metastasis and to train DFNN models via grid search and 5-fold cross-validation (Xu et al., 2022). A complementary breast-cancer prognostic framework uses routine clinical and histopathologic variables together with Bayesian-network risk factor learning and multiple classifiers to predict 5-, 10-, and 15-year metastasis, with validation AUC up to 0.9 (Jiang et al., 2024). At the molecular level, genotype–phenotype integration for metastasis prediction combines CCLE gene expression with a DoRothEA transcription factor–target prior focused on nine metastasis-associated regulators, then constructs personalized gene regulatory networks using PANDA and LIONESS for graph-based prediction (Fu et al., 22 Oct 2025). More generally, CRC radiomics studies treat CT, MRI, PET-CT, and related modalities as quantitative phenotyping sources for lymph-node metastasis, hepatic metastasis, prognosis, and treatment response models (Karami et al., 2024).

4. Interpretability, evidence grounding, and relational reasoning

Interpretability is a defining concern in metastasis AI agents because the target decisions often alter staging, therapy, or surveillance. In liver metastasis assessment, the predictive model is formalized as $f:\mathbb{R}^m \rightarrow \mathbb{R}$, and pixelwise Shapley values are approximated with HyperSHAP to produce visual gradients showing which pixels or regions increased or decreased lesion probability (Hunt et al., 2019). These attributions are exposed through the user interface, where clinicians can request explanations for AI-detected lesions or manually tagged regions.

In pathology, interpretability is often expressed as structured spatial coherence or attention rather than explicit natural-language rationales. NCRF improves patch probability maps by modeling spatial correlations between neighboring patches through pairwise CRF potentials, reducing “salt-and-pepper” false positives and sharpening tumor boundaries (Li et al., 2018). AMD-MIL computes attention heatmaps from agent-mediated aggregation, and those heatmaps overlap strongly with tumor-annotated regions on CAMELYON; the model is reported to localize both macro-metastases and micro-metastases, and the mask denoise mechanism concentrates attention more on cancerous ROIs while reducing attention to normal areas (Ling et al., 2024). In cleaned benchmark settings, the persistent difficulty of ITC classification makes clear that explanation alone is not sufficient; the representation must also capture lesion size and class imbalance (Ling et al., 2024).

Theranostic agents extend interpretability into explicit evidence-grounded reasoning. TheraAgent stores prior cases as $m^{(j)} = (\mathbf{f}^{(j)}, y^{(j)})$, retrieves similar cases from SEA-Mem, and computes an empirical case probability $$P_{\text{case}} = \frac{1}{|\mathcal{S}|}\sum_{s \in \mathcal{S}} y_s$$ over the top-$k$ retrieval set; in parallel, it queries a curated theranostics knowledge base for passages from VISION, TheraP, and related studies (Chen et al., 14 Mar 2026). This makes explanation multi-source: structured features, similar prior cases, and trial evidence. A plausible implication is that metastasis AI agents increasingly treat interpretability as a layered object—pixel or patch saliency, relational structure, and external evidentiary grounding—rather than as a single post hoc visualization.

5. Clinical tasks and reported performance

The literature covers detection, staging, prognostication, response assessment, and intra-operative triage. In pathology, DenseNet169 plus RAdam on the modified PCam dataset achieved AUC-ROC 0.9650 and accuracy 0.980 for patch-level metastatic cancer classification, outperforming VGG19, ResNet34, and ResNet50 in that study (Qiu et al., 2020). NCRF achieved an average FROC score of 0.8096 on the Camelyon16 test set and improved patch-classification accuracy relative to baseline CNNs by explicitly modeling spatial correlations (Li et al., 2018). The breast-cancer pN-stage pipeline achieved a state-of-the-art quadratic weighted kappa of 0.9203 on Camelyon17 (Lee et al., 2018). In weakly supervised WSI classification, AMD-MIL reported CAMELYON-16 performance of ACC 92.9%, AUC 96.4%, and F1 92.7%, and CAMELYON-17 performance of ACC 85.0%, AUC 85.3%, and F1 82.7% (Ling et al., 2024). On the reprocessed Camelyon$768 \times 768$0 benchmark, the best pathology foundation models with advanced MIL reached approximately 92–93% macro AUC and approximately 86–88% accuracy for four-class slide classification, while ITC remained the hardest class (Ling et al., 2024).

In radiology and procedural guidance, the intra-operative CASL prototype outperformed oncologic surgeons in a simulated environment for peritoneal surface metastases, with the reported effect that the AI model would have improved identification of metastases by 5% while reducing unnecessary biopsies by 28% compared to current standard practice (Schnelldorfer et al., 2023). In small-data neuroimaging, deep neuroevolution for neuroblastoma brain metastasis classification converged to 97% test accuracy on a testing set enriched with outside images and 96% accuracy on the outside-only subset, emphasizing robustness to heterogeneous multi-institutional MRI (Purkayastha et al., 2022). In longitudinal MRI after SRS for brain metastases, 12-month lesion-level response prediction reached up to 0.90 AUC using gradient boosting from pre-treatment and first follow-up MRI, and up to 0.88 AUC using graph machine learning (Kuhn et al., 8 Sep 2025).

In structured clinical prediction, iMedBot reports a current best validation AUC of 84.3% for pretrained breast-cancer metastasis models and shows a user-trained example with “Mean Validation ROC (area=0.9350)” (Xu et al., 2022). The breast-cancer “coalitions” framework reported validation AUC up to 0.9 using routine clinical and histopathologic data alone, with DNM_RF-15Year reaching mean-test AUC 0.862 and LASSO-15Year reaching validation AUC 0.901 (Jiang et al., 2024). In metastatic prostate cancer theranostics, TheraAgent achieved 75.7% overall accuracy on 35 real patients and 87.0% on 400 synthetic cases, outperforming MDAgents and MedAgent-Pro by over 20% on the reported benchmark (Chen et al., 14 Mar 2026). These results show that the same umbrella concept can support very different clinical endpoints: lesion detection, size-based classification, long-horizon risk, treatment response, and workflow triage.

6. Limitations, controversies, and future directions

The major limitation is that “Metastasis AI Agent” is not a single standardized technical object. Some systems are proof-of-concept interactive reporting frameworks, some are pathology classifiers, some are web applications, some are internal trainable tokens, and some are large multi-agent reasoning architectures (Hunt et al., 2019, Ling et al., 2024, Xu et al., 2022, Wang et al., 27 Aug 2025). This terminological breadth can obscure important differences in autonomy, evidence use, and validation status.

Several recurring technical constraints remain unresolved. Interactive liver-metastasis assessment was presented as a framework with lesion predictions, Shapley-based explanations, and structured reports, but without full clinical validation or formal user studies, and the model was trained only on liver metastases (Hunt et al., 2019). NCRF was tuned to Camelyon16 at 40× and models only binary tumor-versus-normal patch labels within fixed $768 \times 768$1 neighborhoods, so larger context and other tissue domains require adaptation (Li et al., 2018). The cleaned Camelyon$768 \times 768$2 benchmark demonstrates that label quality and class imbalance materially affect apparent progress; ITC remains a long-tailed, failure-prone class even with better labels and pathology foundation models (Ling et al., 2024). TheraAgent relies on only 35 real patients, uses reports rather than raw images, and is specific to PSMA-positive mCRPC treated with $768 \times 768$3Lu-PSMA-617 (Chen et al., 14 Mar 2026). The Full-Body AI Agent formulation explicitly notes that comprehensive multi-scale data are rare and costly, and that interpretability, scalability, and prospective validation remain open challenges (Wang et al., 27 Aug 2025).

Future directions therefore converge on several themes. One is broader multimodality: combining radiology, pathology, omics, spatial data, clinical records, and longitudinal follow-up so that metastasis is modeled across biological and temporal scales rather than as a static image label (Wang et al., 27 Aug 2025, Karami et al., 2024, Fu et al., 22 Oct 2025). Another is more explicit human–AI collaboration: richer interfaces for lesion review, structured reporting, and evidence display rather than hidden inference (Hunt et al., 2019, Schnelldorfer et al., 2023). A third is stronger agentic memory and cross-case reasoning, where small but high-quality cohorts can be amplified by retrieval, case accumulation, and evidence-grounded synthesis (Chen et al., 14 Mar 2026). This suggests that the mature metastasis AI agent will likely be neither a pure classifier nor a fully autonomous diagnostician, but a modular system that detects, explains, predicts, retrieves evidence, and supports clinician judgment across metastatic disease states.