AI System Uses Multi-Step Reasoning and Self-Checking to Improve Endoscopic Diagnosis

Medical AI for endoscopy faces a critical challenge: traditional systems struggle to match the way doctors actually work—examining images repeatedly, cross-checking findings, and refining conclusions as they gather more evidence. Existing approaches often produce hallucinated observations or let errors compound uncorrected, weakening diagnostic reliability.

A Closed-Loop Framework for Endoscopic Reasoning#

Researchers have proposed EndoCogniAgent, a reasoning system that restructures endoscopic diagnosis as a continuous feedback loop. Rather than making a single pass through an image, the framework cycles through repeated rounds of observation and validation. At each step, a central planner decides what evidence to acquire next. Specialized tools then extract both visual details and semantic information from the endoscopic images, translating them into structured text. A built-in consistency checker evaluates these findings in two ways: whether they align with what's actually visible in the image, and whether they contradict anything the system has already validated. Only evidence that passes both checks gets added to the diagnostic record; conflicting or weak findings are flagged with corrective feedback that guides the planner to seek additional proof.

Testing at Scale Across Multiple Specialties#

The team introduced EndoAgentBench, a comprehensive evaluation dataset drawn from 11 separate endoscopy datasets and containing 6,132 question-answer pairs. The benchmark spans the full diagnostic workflow, from fine-grained perception tasks (identifying small structures in images) to high-level clinical reasoning. On perception tasks, EndoCogniAgent achieved 85.23% accuracy. For reasoning tasks, the system reached a 71.13% clinical acceptance rate—a metric that reflects how often clinicians would find the diagnostic response sound enough to rely on in practice. Independent blinded evaluation by physicians confirmed that the system's responses outperformed comparison baselines.

Why This Matters for Medical AI#

The design directly addresses two persistent failure modes in medical AI: hallucination (inventing evidence not present in images) and uncorrected error cascades (when early mistakes propagate through downstream reasoning). By enforcing multi-round evidence gathering and validation checkpoints, the framework shifts AI diagnosis closer to human clinical practice, where verification and re-examination are central to confidence.