1029 - SMART: A Multi-Center Clinical Practice Driven by an AI-Enhanced Human-in-the-Loop Workflow in Breast Cancer

Purpose/Objective(s): To evaluate the clinical feasibility of an AI-enhanced human-in-the-loop workflow (SMART, Single-session Multi-step AI-interactive Radiation Therapy), which enables one-step radiation therapy from simulation to treatment for breast cancer patients on a CT-linac integrated system, and to appraise its contouring and planning quality, efficiency, and clinical implementation.

Materials/Methods: Totally 150 consecutive breast cancer patients diagnosed with N+ stages (111 breast-conserving [BCS], 39 mastectomy [TM]) were prospectively enrolled, and 12 patients (6 BCS, 6 TM) from an external center were retrospectively included. SMART was structured around two deep learning-based modules: (1) Auto-contouring: nnU-Net models trained on an institutional dataset generated initial target volumes (CTVp and CTVnd), and OARs. (2) Auto-planning: A CAD-UNet for dose prediction was combined with an automated optimization engine, configured per institutional criteria (50Gy/25f) to generate deliverable treatment plans. After CT simulation, initial contours from module 1 were reviewed/adjusted by radiation oncologists, then forwarded to module 2 for plan generation. The resulting plan underwent a focused clinical team oversight before same-session delivery. We assessed contouring accuracy (DSC), plan quality (DVH indices) and recorded the total end-to-end session time along with the duration of each individual step.

Results: SMART demonstrated high performance. Auto-contouring achieved excellent DSC for primary targets: 0.99±0.01, 0.98±0.03 and 0.79±0.15 for CTVp, CTVsc and CTVim in BCS cases, and 0.92±0.07, 0.92±0.05 and 0.82±0.10 in TM cases, respectively. OAR contours were consistently compliant with clinical standards and approved as final without modifications. All plans satisfied clinical coverage criteria (PTVs V95>95%) and met all dose constraints for all OARs. Notably, 52.7% (79/159) of plans achieved excellent dosimetric standards (PTVs V95>97%, ipsilateral lung V5<50%, and heart mean dose <8/5 Gy for left/right cases). A subset of 49 plans was directly accepted without any manual adjustment. The complete simulation-to-treatment workflow was accomplished in average of 24.5±6.4 min:

Conclusion: The study validates the practical viability of SMART workflow, enabling a consistent ~25 min simulation-to-treatment timeline while preserving high contouring accuracy and plan quality. By merging AI-automation and clinical oversight, SMART establishes human-AI collaboration, boosting patient convenience and clinical throughput.