Technical Summary: The Path to Self-Evolving Clinical Systems

Problem Statement

Current medical artificial intelligence, particularly foundation models in medical imaging, remains fundamentally passive. These systems typically map inputs to outputs in a single forward pass, lacking the capacity for iterative reasoning, tool invocation, or contextual adaptation. While large language models (LLMs) and vision-language models (VLMs) have advanced, they often fail to address the realities of clinical deployment: partial observability (where the agent never sees the true patient condition directly), the need for multi-step workflows, and the requirement for actions to be auditable and reversible.

Existing surveys often organize medical agents by cognitive modules or tools, treating evaluation as a downstream appendix. This approach obscures the critical question of "real-world readiness." The field lacks a unified framework to understand how heterogeneous systems scale from simple assistance to full autonomy, and how they can improve through interaction with clinical environments rather than solely through parameter growth.

Methodology and Framework

This survey adopts a "real environment first" organizing principle. Instead of starting with model capabilities, it begins with the constraints of clinical deployment. The authors formalize medical agents as a sequential decision process under partial observability, defined by six interacting cognitive modules: Perception, Reasoning, Planning, Memory, Tool Use, and Action/Reflection.

To organize the diverse literature, the paper introduces a Scaling Spine comprising three orthogonal, complementary axes:

1. Framework Scaling ($N_{topo}$): Scales the structural richness of agent topologies, moving from single tool-augmented agents to role-based teams and dynamic orchestration networks.
2. Capability Scaling ($D_{loop}$): Scales the depth of the cognitive loop (perceive-reason-plan-act-reflect) around a fixed backbone, emphasizing the engineering of the runtime harness over raw parameter count.
3. Environment Scaling ($|E|$): Scales the reachable environment by enriching the tools, data sources, and interfaces (e.g., PACS, EHR, FHIR) available to the agent. The authors argue this is the most immediately actionable lever for medical agents.

The survey also proposes a Three-Level Autonomy Taxonomy:

• Level 1 (Assisted): The agent suggests findings under direct clinician control.
• Level 2 (Cooperative): The agent decomposes tasks and invokes tools within a human-in-the-loop framework.
• Level 3 (Fully Autonomous): The agent executes end-to-end workflows with exception-based human escalation.

Key Contributions

1. Real Environment-Centric Organization: The survey reorganizes the field by starting with clinical scenarios, contamination-resistant benchmarks, and interactive training environments ("Clinical Gyms"), rather than leading with generic capability taxonomies.
2. Formal Definition and Taxonomy: It provides a formal definition of medical agents as partial-observability decision processes and unifies the literature under a three-level autonomy taxonomy.
3. The Scaling Spine: It introduces a global map (the Scaling Spine) that threads the entire field through framework, capability, and environment scaling, distinguishing these from traditional parameter and test-time compute scaling.
4. Self-Evolution as a Frontier: It elevates "self-evolving systems" from a niche capability to the destination of the scaling spine, synthesizing lessons from general domain self-improving agents and transferring them to medical imaging.
5. Comprehensive Benchmark and Risk Analysis: It catalogs over 300 references, including 2025–2026 advances, and reviews applications across radiology, pathology, ophthalmology, and oncology. Crucially, it analyzes risks including hallucination, cascade failures in multi-agent systems, and governance issues (fairness, privacy, liability).

Results and Findings

• Benchmark Maturity: The field is shifting from Level 1 benchmarks (static, single-capability tasks like VQA) toward Level 3 benchmarks (system-level, end-to-end workflows in real or simulated EHRs). Current systems are competent perceivers but immature operators, often failing when required to sustain multi-turn dialogues or manage long-horizon workflows.
• Training Paradigms: The survey traces a ladder of training strategies from prompting and supervised fine-tuning (SFT) to preference optimization (DPO) and reinforcement learning with verifiable rewards (RLVR). It highlights self-play in simulated clinical environments (e.g., Agent Hospital, MedAgentSim) as a critical step toward autonomous learning where agents improve through experience without weight updates.
• Application Landscape:
  • Radiology: The most mature specialty, with agents reaching multi-agent collaboration and autonomous workflow management (e.g., RadAgent, MedRAX).
  • Pathology: Agents are emerging to handle gigapixel whole-slide images via zoom-and-focus strategies and multi-agent search.
  • Ophthalmology & Neurology: Progress is gated by trustworthiness and hallucination detection rather than raw accuracy; recent work focuses on traceable reasoning and fairness.
  • Cross-Specialty: Systems are moving toward generality via adaptive collaboration topologies (e.g., MDAgents) and evidence-grounded reasoning.
• Deployment Barriers: The binding constraints for deployment are rarely model accuracy but rather interoperability (integration with DICOM/FHIR), productization (ambient documentation vs. autonomous diagnosis), and regulatory compliance (reconciling self-evolving models with static approval frameworks like the FDA's Predetermined Change Control Plan).

Significance and Claims

The paper claims to provide the first roadmap that explicitly couples a disclosed retrieval protocol with an environment-centered scaling axis. Its primary significance lies in shifting the narrative from "how powerful the model is" to "how the agent interacts with the clinical ecosystem."

The authors argue that environment scaling—enriching the tool and data environment—is the most immediate path to advancing medical agent capabilities, often yielding more readiness per unit effort than additional finetuning. Furthermore, they position clinical self-evolution (the ability of an agent to improve through environment interaction) as an aspirational frontier rather than a solved problem. The survey concludes that while self-evolving medical agents are not yet a solved capability, the convergence of multimodal foundation models, standardized tool integration, and environment scaling offers a credible path toward trustworthy, self-improving systems ready for real clinical settings. The work serves as a call to action for researchers, clinicians, and regulators to collaborate on the safety, equity, and governance of these emerging systems.