Purpose: The standard of care for multiple brain metastases (BrM) is shifting from whole brain radiotherapy to stereotactic radiosurgery (SRS). The radiation oncologist must contour every gross tumor volume (GTV) for treatment planning—a time-consuming task. We present a streamlined clinical process that leverages artificial intelligence (AI) segmentations to aid in contouring GTVs at post-contrast MR and CT images.
Methods: A workflow was created in the treatment planning system (TPS) for exporting co-registered CT simulation and MR scans to a network listener server. The server executed AI inference on a GPU cluster using our published multi-stage 3D CNN model, then returned to the TPS a structure set containing AI GTVs in about 5 minutes. Physicians were then able to load the patient in the TPS, modify the AI GTVs as needed, and approve volumes for treatment. Strict naming conventions were followed. This process underwent a pilot test with 44 patients. The accuracy was quantified by detection sensitivity, Dice coefficient, and Hausdorff distance. Time savings were estimated based on GTV size and the amount that was edited.
Results: Averaged per patient, 4 ± 4 lesions were prescribed RT, and there were 2 ± 2 false positive AI GTVs. AI had good sensitivity for lesions with prescription (92%, 152/170). GTV segmentations were accurate, with median Dice coefficient 0.87 (IQR [0.80, 0.95]) and Hausdorff distance 2.0 mm (IQR [1.4, 3.1]). The median size of edits required was 0.10 mL (IQR [0.04, 0.29]). Physician time savings were approximately 15 minutes/patient.
Conclusion: This process lays the groundwork to apply AI GTVs in clinical treatment planning for multiple BrM’s. Providing clinicians with assistance from AI can streamline the volumes task and shorten the timeline from simulation to treatment.
Stereotactic Radiosurgery, Segmentation, Brain