Clinical Validation of an AI-Driven Detection Assistant for Gamma Knife Radiosurgery: Improving Workflow Confidence and Target Identification
Abstract
Purpose
To evaluate the clinical feasibility of an in-house AI-driven detection tool for brain metastases in Gamma Knife radiosurgery. Small or interval metastases are prone to oversight during initial radiologic review yet require identification during the pre-treatment workflow. This study aims to deploy an in-house deep learning model specifically optimized to flag these subtle, small-volume lesions, acting as a safety filter to ensure comprehensive treatment delivery.
Methods
T1-post MRI sequence data from 368 patients were utilized to train a custom detection model using the nnU-Net framework, with a separate 93-patient dataset reserved for validation. To address the clinical challenge of identifying small lesions (105 mm³, sensitivity 0.96) and medium tumors (0.88), but limited performance for small tumors (0.57). Implementation of the TopK loss function marginally improved overall DSC to 0.83 but yielded a marked improvement in small lesion detection, increasing sensitivity to 0.63 (>10% improvement). Sensitivity for medium (0.90) and large (0.96) tumors remained high.
Conclusion
This study demonstrates the potential of an optimized in-house AI model to serve as an automated screening tool for Gamma Knife planning. While global DSC remained stable, the integration of TopK loss significantly enhanced sensitivity to small, easily overlooked tumors. By acting as a supplementary "second read," this tool streamlines metastasis identification, minimizing the risk of missed targets and ensuring the high precision required for effective patient outcomes.