Dual-Tier OncoAgent Hits 0.842 Zero-Shot DSC in Radiotherapy

Research detailing OncoAgent, a specialized multi-agent framework for oncology clinical decision support, was published on May 9, 2026, following the AMD Developer Hackathon. The framework automates the delineation of clinical target volumes by linking 3D medical imaging directly to unstructured textual clinical guidelines. This produces a training-free, zero-shot pipeline that adapts to clinical standard updates without requiring dataset rebuilding or model retraining. By explicitly linking delineations to text, the system addresses the interpretability gap in radiotherapy, where clinicians previously struggled to audit why an AI selected a specific margin.

Dual-Tier Agent Architecture

Rigid deep learning models rely on pre-annotated datasets that age poorly as clinical consensus shifts. OncoAgent solves this by adopting hierarchical multi-agent coordination patterns that split rule extraction from spatial execution.

The Tier 1 Guideline Processing Agent continuously parses clinical text, such as NCCN or ESTRO guidelines. It extracts precise anatomical rules, spatial constraints, and margin requirements from the unstructured documents. The Tier 2 Execution and Delineation Agent then projects these extracted rules onto 3D medical imaging, evaluating CT and MRI scans to generate target contours. This decoupled approach creates a fully transparent audit trail, allowing medical staff to trace exactly which sentence in a guideline document produced a specific boundary decision.

Zero-Shot Benchmark Results

The framework was evaluated on esophageal cancer cases to test its capacity for zero-shot clinical target volume (CTV) and planning target volume (PTV) generation.

Target Volume	Zero-Shot Dice Similarity Coefficient (DSC)
Clinical Target Volume (CTV)	0.842
Planning Target Volume (PTV)	0.880

These segmentation scores are highly comparable to the fully supervised nnU-Net baseline, the current standard in medical image segmentation. The framework also generalized across alternative esophageal guidelines and entirely different anatomical sites, including the prostate, with zero additional training steps. In a blinded clinical evaluation, oncologists preferred the OncoAgent contours over supervised baselines, rating the framework higher for guideline compliance, clinical acceptability, and reduced modification effort.

Inference Hardware and Privacy

Processing high-resolution 3D medical imaging alongside complex agent routing requires substantial memory bandwidth. The system relies on AMD MI300X using ROCm deployed via the AMD Developer Cloud to sustain high-throughput 3D image processing.

Clinical deployments face strict data sovereignty requirements. Developed under the hackathon’s “CLOAK-AI” and “AgentOps” themes, the framework isolates the execution environment. Patient imaging data remains entirely within the clinical perimeter while the guideline processing agents coordinate the extraction logic externally.

If you build automated diagnostic pipelines, decoupling rule extraction from spatial execution provides a reliable architectural blueprint for regulatory compliance. Keeping raw imaging payloads strictly separated from external reasoning models allows you to deploy advanced logic while preserving patient data boundaries.