335 - AI as a Safety Net: Clinical and Dosimetric Impact of Intercepting Undetected Brain Metastases in SRS/SRT
Presenter(s)
M. Kinz1, A. A. Aizer1, R. Rahman2, S. Tanguturi2, R. H. Mak1, V. Nappady Joy3, J. Hesser4, C. V. Guthier5, K. Singhrao1, and A. Sudhyadhom1; 1Department of Radiation Oncology, Brigham and Women’s Hospital, Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA, 2Department of Radiation Oncology, Mass General Brigham Cancer Institute, Boston, MA, 3Siemens Healthineers, Forchcheim, Bayern, Germany, 4Mannheim Institute for Intelligent Systems in Medicine, Medical Faculty Mannheim; Interdisciplinary Center for Scientific Computing; Central Institute for Computer Engineering; CZS Heidelberg Center for Model Based Ai; Heidelberg University, Heidelberg, Germany, 5Department of Radiation Oncology, Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA
Purpose/Objective(s):
Accurate identification of brain metastases is critical for effective stereotactic radiosurgery/-therapy (SRS/SRT), yet small or subtle lesions can occasionally go undetected, requiring subsequent salvage treatments. This leads to increased resource utilization, potential morbidity such as radionecrosis, overlapping radiation fields, compromised clinical trial eligibility if intracranial control is required, and a loss of patient time. To mitigate this, we retrospectively investigated a deep learning auto-contouring prototype as a clinical "safety net." We hypothesized that early AI-guided intervention could identify undetected metastases, avert tumor growth, achieve dosimetric tissue sparing, and reduce the risk of radionecrosis.Materials/Methods:
We retrospectively analyzed 436 brain metastasis patients treated with SRS/SRT (11/21-04/25). Initial AI false positives (FPs) were matched against subsequent courses to classify Retrospectively Identified Metastases (RIMs, lesions not marked at Time 1 but treated at Time 2), evaluating RIM incidence and interval volumetric growth as clinical endpoints. For a proof-of-concept cohort (n=5), a retrospective analysis isolated the dosimetric cost of delayed detection. This counterfactual analysis evaluated cumulative brain V12Gy and estimated radionecrosis risk as dosimetric endpoints, comparing the clinical reality against a simulated early-intervention timeline (re-planning Course 1 with RIM, Course 2 without).Results:
AI sensitivity for lesion detection was 90%. 42 RIMs were found in 27 patients, representing 23% of the 117 multi-course patients and 6% of the total cohort. After subtracting the RIM, the median number of persistent, untreated FPs per patient was reduced to 0 (0-10), indicating high specificity of the AI. The cohort had a median burden of 4 treated clinical lesions (1-20) per initial course. Over a median 121-day delay (55-1013), RIMs grew by a median 538% (0.02 to 0.15 cc; 3.5 to 6.5 mm). In the simulated re-planning cohort, early RIM inclusion yielded a median absolute brain V12Gy reduction of 5.1 cc (1.8-19.9 cc) and a 21% (2-52%) relative reduction. This translated to an estimated median absolute radionecrosis risk reduction of 12% (5-20%). In one notable case, early AI detection could have prevented a lesion from expanding near the visual pathway, sparing the optic nerve and chiasm and eliminating the need for a separate subsequent SRS plan delivery.Conclusion:
By intercepting undetected metastases, an AI auto-contouring safety net could potentially avert tumor growth, spare healthy brain tissue, mitigate radionecrosis risk, protect critical structures, and reduce the need for complex salvage radiation. Larger-scale dosimetric validation, clinical adjudication of RIMs to determine intentionality, and a prospective trial are warranted.