RTAD-11 DOSIMETRIC AND GEOMETRIC IMPACT OF AN AUTOMATED TUMOR DELINEATION TOOL FOR INTRACRANIAL METASTASES

Abstract

Stereotactic radiosurgery (SRS) is the standard of care for the majority of patients with brain metastases. Treatment of increasing numbers of metastases highlights the need for tools enabling consistent tumor identification and delineation. This study aimed to evaluate the performance of a novel nn-Unet AI tool on consistency in tumor definition and radiation plan coverage than standard manual approaches. In this prospective multi-institutional study across 3 institutions, AI pre-contours for 149 detected lesions from 20 patients with 163 lesions were generated on contrast enhanced MRI. Three expert physicians independently also delineated these metastases. Three months later, to assess geometric variation, physicians modified a randomized mix of their own and AI pre-contours. For radiation dosimetric analysis, seven (three physician, AI, and three physician-edited AI) contour sets per lesion were evaluated. SRS treatment plans were generated using both linear accelerator–based and Gamma Knife platforms. Tumor dose metrics were categorized and classified by observer and lesion size. Physician-edited AI-generated pre-contours showed greater geometric consistency (Dice 0.89 vs 0.72, p < 0.001 and 95th percentile Hausdorff distance 0.82 mm vs 1.54 mm, p = 0.005) over physician-only tumor delineation. Tumor coverage dosimetric metrics were more consistent using AI-generated vs physician expert-generated volumes, for both linear accelerator and Gamma Knife plans. For tumor volume strata of 0–0.1cc, 0.1–1cc, and >1cc, greater consistency in D100% (p < 0.001, <0.001, <0.001), V100% (p < 0.001,=0.0003,=0.41), and Dmin (p = 0.00002,=0.002,=0.015) were observed, respectively.  Compared to physician expert brain metastasis delineation without AI assistance, AI-based delineation provided to physicians as pre-contours improved consistency as well as dosimetric metrics known to be associated with local control. These findings suggest potential value in integrating select automated delineation tools to assist in high quality radiosurgery treatment planning with potential application in patient care and clinical trials.