AI-DRIVEN MR IMAGE FEATURES VERSUS RANO-BM CRITERIA IN DISTINGUISHING RECURRENT BRAIN METASTASES FROM RADIATION TREATMENT EFFECT: A COMPARATIVE, MULTIINSTITUTIONAL STUDY
Recommended Citation
Um H, Ismail M, Hill V, Puri S, Yu J, Lu L, Nayate A, Rogers LR, Prasanna P, Bardhan M, Li C, Basree M, Baschnagel A, McMillan A, Bhatia A, Ahluwalia M, Veronesi M, Tiwari P. AI-DRIVEN MR IMAGE FEATURES VERSUS RANO-BM CRITERIA IN DISTINGUISHING RECURRENT BRAIN METASTASES FROM RADIATION TREATMENT EFFECT: A COMPARATIVE, MULTIINSTITUTIONAL STUDY. Neuro-Oncology Advances 2024; 6(Supplement_1):i10.
Document Type
Conference Proceeding
Publication Date
8-1-2024
Publication Title
Neuro-Oncology Advances
Abstract
A significant challenge in brain metastases (BM) management is distinguishing radiation-induced treatment effect (TE) from tumor recurrence (TR). TE mimics the appearance of TR on follow-up MRI, making radiographic diagnosis unreliable. The standardized Response Assessment in Neuro-Oncology for brain metastases (RANO-BM) is suboptimal due to high inter-reader variability. We compared the performance of artificial intelligence (AI)-driven MRI features with that of RANO-BM criteria in differentiating TE from TR. We hypothesize AI-features from routine MRI can capture the pathophysiologic differences between TE and TR, occult on structural MRI and hence overlooked in standard-of-care evaluation. A total of 261 lesions with pathologically-confirmed diagnoses in 189 patients were retrospectively analyzed. 201 lesions (111 TR,90 TE) from Cleveland Clinic and University Hospitals, Cleveland were used for training a machine learning model. 60 lesions (33 TR,27 TE) from University of Wisconsin- Madison were used for model testing. MRI (Gd-T1w, T2w, FLAIR) were preprocessed, and lesions were expertly segmented into enhancing lesion, edema, and necrosis. 856 texture features were extracted from each sub-compartment, and a random forest classifier was employed for 3-fold cross-validation. Top-performing features and RANO-BM criteria were evaluated on the test set. Results show T1 features from edema were most discriminatory in differentiating TR from TE (training-AUC=0.86, testaccuracy= 71.7%, test-sensitivity=78.8%). Using RANO-BM, 9 cases were excluded due to lack of longitudinal imaging to estimate lesion growth. Additionally, since no lesions decreased in sum of longest diameter, none were classified as partial response while the remaining 51 cases were classified as stable disease (n=14, (8 TR,6 TE)) or TR (n=37, accuracy=54.1%). Interestingly, 78.6% of the stable lesions were accurately classified using our AI-model as TE or TR, missing only 3 cases (2 TR,1 TE). Our results suggest AI-driven models on clinical MRI scans may reliably distinguish TR from TE, demonstrating potential utility in clinical practice.
Volume
6
Issue
Supplement_1
First Page
i10
