Evaluation of an artificial intelligence-based tool for brain metastasis auto-detection: A comparison with physician review
Sophia Shah, Haisong Liu, Zhenghao Xiao, Emily Meinert, Kelsey Carey, Wentao Wang, Yingxuan Chen, Kevin D Judy, Kiran Talekar, James J Evans, Rui Feng and Wenyin Shi
Purpose: To evaluate an artificial intelligence–based automatic lesion detection (ALD) tool for identifying brain metastases compared with experienced physicians.
Methods: In this retrospective study, 41 patients treated with stereotactic radiosurgery (2016–2023) were analyzed. Brain metastases were contoured on three-dimensional contrast-enhanced T1-weighted magnetic resonance imaging and compared with lesions detected by the ALD tool. An independent three-physician panel established the reference standard.
Results: Physician contoured 223 lesions, and ALD identified 236, with 203 concordant. Among 53 discrepant lesions, 30 were confirmed as true (21 ALD only, 9 physician only). Overall, ALD correctly detected 224 true lesions, with 5 false positives and 9 misses [sensitivity 96.1%, positive predictive value (PPV) 97.8%]. Physician contouring identified 212 true lesions, with 7 false positives and 21 misses (sensitivity 91.0%, PPV 96.8%).
Conclusion: The ALD tool demonstrated high sensitivity and PPV, potentially improving efficiency in treatment planning. However, false positives and missed lesions highlight the need for physician oversight and prospective validation.
Keywords: AI, brain metastasis, auto-detection, radiosurgery
Full Text (IP)
Purchase Article (PDF)