This study develops a collaborative framework in which several agents analyze a medical question from different professional perspectives, discuss the evidence, vote on the proposed answer, and revise unresolved questions up to three times, showing that structured collaboration can make medical artificial intelligence workflows more transparent.
Abstract
Medical artificial intelligence systems often rely on a single model, which may give inconsistent answers and does not reflect the team-based nature of clinical consultation. This study develops a collaborative framework in which several agents analyze a medical question from different professional perspectives, discuss the evidence, vote on the proposed answer, and revise unresolved questions up to three times. Each step and stopping condition is recorded in a time-ordered workflow so that the decision process can be reviewed. The framework was evaluated on three regional versions of a medical examination dataset and three medical image question-answering datasets. It achieved an accuracy of 80.1% on the Mainland examination dataset and produced the highest reported results among the evaluated systems on ADAM. Stronger image models performed better on ACRIMA and Covid CT, showing that collaboration remains limited by the underlying visual model. Additional model-based quality scores are reported only as exploratory results because they were not validated by an independent model or human experts. These findings show that structured collaboration can make medical artificial intelligence workflows more transparent.
Complex clinical reasoning requires models to update diagnostic hypotheses as new evidence emerges and to coordinate different medical specialities under limited consultation resources. Existing LLM-based clinical reasoning systems typically perform single-pass prediction or rely on fixed multi-agent workflows, making expert participation either static or unnecessarily exhaustive. We propose Sparse Multi-Stage Expert-Agent Routing, a language-based clinical reasoning framework that models diagnosis as a stage-wise routing process. Given progressively available clinical evidence derived from multiple modalities, the framework maintains an evolving case state and adaptively activates a sparse set of medical expert agents, supported by expert-specific memory across stages. To evaluate free-text diagnostic conclusions beyond surface similarity, we further introduce ClinFEScore, a fact-aware semantic evaluation protocol for clinical reasoning outputs. On reconstructed multi-stage cases from MAC and AgentClinic-NEJM, our framework reduces the average number of activated experts from 17.0 to 3.0 whilst maintaining strong fact-level diagnostic quality. On 200 real-world hospital MDT cases, ClinFEScore correlates strongly with clinician judgements (Spearman's $\rho=0.81$; Pearson's $r=0.87$), whilst our method achieves 91.5\% clinician-verified diagnostic accuracy with approximately five expert-agent/LLM calls per case. These results support sparse stage-wise coordination as an efficient and clinically relevant approach to LLM-based clinical reasoning.
Sike Xiang, Shuang Chen, Qianpeng Sun et al.· 0 citations
Multimodal Artificial Intelligence (AI) models-integrating diverse data such as imaging and clinical records-are advancing rapidly in healthcare, yet a significant disconnection persists between these complex predictive architectures and the explainable AI (XAI) techniques used to interpret them. We conducted a scoping review over 4 bibliographic databases to investigate the use of explainability methods in cross-modal medical AI studies. From 82 included studies, we found that the landscape remains dominated by independent feature attribution (assigning importance scores to individual modality in isolation), with the majority of studies relying on post-hoc methods (applied after a model decision is reached) that treat the model as a 'black box'. While emerging trends like visual grounding (linking textual justifications directly to specific image regions) and model reasoning show promise, a critical gap remains in explaining the underlying reasoning process. Standardised evaluation is missing in the majority of studies relying solely on qualitative measures. Only a minority of studies achieve good reproducibility with public codebase. We provide suggestions for the field to transition from individual and post-hoc XAIs toward intrinsically explainable designs where the reasoning logic is built directly into the model architecture to ensure that AI outputs align with human-centric clinical workflows and applications.
Kai Hu, Xing-Yue Fu, Yupeng Zhang et al.· npj Digital Medicine· 0 citations
Although promising, LLM-based systems are not yet reliable enough for autonomous medical diagnosis, and multiple recommendations for future research are contained to ensure a high level of safety, transparency, and clinical applicability for LLMs and other AI/ML-related technologies and devices.
M. U. K. Gunawardhna, Pirunthavi Wijikumar, D. Weerasinghe· Sri Lankan Journal of Applie...· 0 citations
Recent advances in artificial intelligence have significantly improved performance in medical imaging tasks such as segmentation, quantification, and report generation. However, most existing solutions operate as static pipelines with limited adaptability, quality assurance, and workflow-level reasoning. In this work, we present ARGUS, an agentic framework for multimodal medical image analysis that coordinates specialized agents within a unified architecture. An Orchestrator Agent interprets user requests, identifies the imaging modality, and assembles task-specific execution plans by selectively engaging processing, quantification, verification, knowledge retrieval, and reporting agents. This enables context-aware decision-making and dynamic workflow reconfiguration based on intermediate findings and runtime conditions. A key feature of ARGUS is its ability to supervise and contextualize analytical processes. The Verification Agent performs quality control by assessing intermediate artifacts against task-specific criteria, while the Knowledge Retrieval Agent enriches quantitative findings with evidence from the biomedical literature and established physiological reference ranges. Together, these components promote transparency, support automated error detection, and reduce the risk of propagating unreliable information through downstream stages. The framework was evaluated across three imaging domains: radiology (MRI), pathology (hematopathology), and ophthalmology (OCT). Quantitative evaluation demonstrated strong agreement between ARGUS and reference standards across pathology, OCT, and MRI tasks, achieving a cell-counting bias of −0.182 cells (MAE = 0.727), a full retinal thickness bias of −31.30μm (MAE = 37.63 μm), and MRI volumetric errors below 3 mL, while also achieving closer agreement with reference measurements than the evaluated general-purpose and domain-specific baseline systems. These results demonstrate the feasibility and potential value of agent-based orchestration for enabling adaptive, validated, and interpretable multimodal imaging workflows while providing a scalable foundation for complex multi-step clinical analysis.
Hoda Helmy, Chaima Ben Rabah, Ahmed Serag· Applied Informatics· 0 citations
Multi-image medical VQA is not merely a prompt-length problem; it is a fundamental challenge of agentic decision-making. Medical vision-language agents must aggregate evidence across ordered images, remain robust to answer-order perturbations, and avoid overfitting to noisy search-time feedback. We study MedFrameQA through a controlled comparison of five inference-time agentic strategies, optimized using the same high-budget ShinkaEvolve configuration and evaluated on a reproducible internal frozen split (1,331 evolution, 665 holdout, 855 final test). Across five independent repeated runs, the strongest method emerges as the simplest robust aggregator: the \textbf{order-vote} policy achieves $57.89 \pm 0.65\%$ final-test accuracy, significantly outperforming the fixed baseline ($52.73 \pm 0.42\%$) and the more complex, albeit brittle, order-rerank variant ($55.79 \pm 0.43\%$). Paired bootstrap analysis confirms these significant gains. Extending the evolutionary search budget from 50 to 100 generations yields no generalization benefit: while holdout performance marginally increases, final-test accuracy drops from $57.89\%$ to $56.02\%$. Our findings suggest that for multi-image medical reasoning, defining the correct agentic decision rule is substantially more impactful than expanding the optimization search budget.
An adaptive memory and reflection (AMR) agentic system, a multi-agent framework in which specialized agents use dedicated memory and reflection-based feedback to retrieve relevant prior cases and improve subsequent reasoning.
P. Murugesan, Luoxiao Yang, Xueli Chen et al.· 0 citations