Radiation Therapy Target Volume Delineation Automation

The workflow automates the repetitive, manual contouring of tumor volumes and critical structures on CT/MRI/PET scans, a bottleneck that consumes 1-3 hours per case. Savings come from direct labor reduction for dosimetrists and physicists, plus accelerated treatment initiation which improves patient throughput and facility revenue. Implementation requires a multi-agent pipeline: a pre-processing agent standardizes DICOM data, a primary segmentation model (e.g., nnU-Net ensemble) contours volumes, and a validation agent scores confidence and checks anatomical plausibility before any data is pushed downstream.

Integration occurs via the TPS API or DICOM-RT export, injecting contours directly into the planning workspace. A dedicated physician review interface displays AI contours with uncertainty maps, enabling efficient edits. Controls are critical: all outputs route through a mandatory review gate; every edit logs to an audit trail for model improvement and compliance. Rollout requires phased validation by disease site (e.g., prostate first, then H&N) to manage clinical risk and build trust in the automated contours before full deployment.

This workflow automates the repetitive, time-consuming manual contouring required in treatment planning systems (TPS) like Varian Eclipse or RayStation. By deploying specialized segmentation models for different anatomical sites, the system ingests simulation CT and fused MRI/PET scans to generate initial contours. The operational upside comes from reducing planner contouring time by 60-70%, directly accelerating the critical path from simulation to first treatment, which improves machine utilization and patient throughput in high-volume oncology centers.

Implementation requires deep integration with the TPS DICOM-RT interface and a physician-facing web console for efficient review and edits. The orchestrator, built on LangGraph, manages the entire pipeline, including confidence scoring and exception routing. Critical controls include an audit trail for all physician overrides to support model retraining and clinical governance, and real-time monitoring of contour quality and processing latency to ensure the workflow meets clinical SLAs without compromising safety.

Phase 1 establishes a secure, human-in-the-loop workflow. The orchestrator ingests DICOM-RT data from the Treatment Planning System (TPS) like RayStation or Eclipse, triggering a segmentation agent to generate initial Gross Tumor Volume (GTV) and Organs-At-Risk (OAR) contours. These are immediately routed to a physician review interface integrated with the TPS, where edits are logged. This initial phase delivers 20-30% time savings for dosimetrists and planners while creating a validated dataset for model refinement and establishing safety protocols.

Phase 2 scales the workflow across multiple centers and introduces automated quality scoring. A confidence-scoring agent evaluates each contour using geometric and probabilistic metrics, triaging low-confidence cases for prioritized review. Integration expands to the Oncology Information System (OIS) for seamless patient context. Phase 3 introduces auto-approval for high-confidence OAR contours, governed by strict confidence thresholds and audit logs, targeting a 50% reduction in manual contouring time. Continuous monitoring for model drift and protocol updates ensures the system adapts to new clinical evidence.

Automated target volume delineation directly reduces planner contouring labor by 60-80% and accelerates simulation-to-treatment timelines, but requires ironclad controls. The workflow must integrate with the Treatment Planning System (TPS) via DICOM-RT, enforce mandatory physician review and edit via a dedicated web interface, and log all overrides to an audit database. A confidence-scoring agent flags low-quality contours for prioritized review, while a rules engine validates anatomical plausibility against clinical protocols before routing to the physician's worklist. This architecture shifts planner effort from manual drawing to quality oversight.

Phased rollout begins with a shadow mode, comparing AI contours to historical clinical contours without impacting treatment. Phase two introduces automation for organs-at-risk only, a lower-risk starting point that builds clinician trust. The final phase activates gross tumor volume automation, initially limited to specific disease sites like prostate. Each phase is governed by a weekly review of key metrics: contour acceptance rate, average edit time, and inter-observer variability. This controlled, data-driven approach mitigates risk while capturing the labor savings and timeline compression that justify the investment.

This workflow automates the most time-consuming step in radiation planning: manually contouring target volumes and critical organs on CT/MRI/PET scans. By deploying specialized segmentation agents, it eliminates 60-90 minutes of repetitive planner work per case, directly accelerating patient throughput. The business case is clear: reducing the simulation-to-treatment planning cycle from days to hours improves machine utilization, increases patient capacity, and lowers labor cost per plan. Implementation requires tight integration with the Treatment Planning System (TPS) like Varian Eclipse or RayStation, using DICOM RT for data exchange and a dedicated orchestrator to manage model inference, quality gates, and physician review routing.

Architecturally, the orchestrator manages a pool of agents—specialized models for different anatomies and modalities—triggered by a DICOM study arrival from the simulator. Each contour undergoes automated confidence scoring and geometric validation; high-confidence results are pushed directly into the TPS via its API, while uncertain contours are routed to a web-based review interface integrated with the physician's workflow. Critical controls include mandatory physician sign-off on all target volumes, a full audit trail of all AI-generated and human-edited contours for regulatory compliance, and continuous performance monitoring against a gold-standard dataset to detect model drift. Rollout is typically phased, starting with OAR contouring to build trust before automating more complex GTV delineation.