Abstract TP161: Machine Learning to Glean Characteristics of Quantitative Susceptibility Maps in Cerebral Cavernous Angiomas with Symptomatic Hemorrhage

Introduction and Hypothesis: New bleeding in cerebral cavernous malformations (CCM) heralds increased risk of future hemorrhage for several years, yet conventional imaging only detects new bleeding that occurred in the prior weeks. A biomarker of hemorrhage could help identifying high risk lesions. An increase of mean lesional quantitative susceptibility mapping (QSM) ≥6% on MRI has been adjudicated as reflecting new bleeding in CCM during longitudinal follow-up. However, mean lesional QSM from a single acquisition could not diagnose or prognosticate a bleed. We hypothesize that machine learning (ML) may identify diagnostic and prognostic features of bleeding within QSM maps at a single point in time. Material and Methods: Two hundred and sixty-five QSM maps of CCM lesions were acquired in 120 patients enrolled in National Institute of Health (NIH) multisite trial readiness project (NCT03652181). Each map was classified (Yes/No) in association with symptomatic hemorrhage (SH) and/or biomarker event with QSM increase ≥6% in the prior (diagnostic association) and subsequent (prognostic association) year. Twenty-eight features were extracted including 14 texture, 5 first-order statistical, as well as 3 size, shape, and morphological. Five-fold cross-validation was conducted on a support-vector machine (SVM) with linear stepwise kernel for both diagnostic and prognostic associations. Performance of individual features and composite classifiers was evaluated using student t-test ( p <0.05) with Bonferroni-correction and receiver operating characteristic (ROC) analysis, area under the curve (AUC). Results: Lesions with SH in the prior year had lower average values for sum variance (AUC=0.79, p<0.001), variance (AUC=0.79, p<0.001), and correlation (AUC=0.71, p=0.004) as compared to lesions without SH. The SVM classification method tasked with distinguishing lesions with mean QSM increase ≥6% in the prior year included recurrent selection of sum average, mean, sphericity and margin sharpness, yielding an AUC of 0.61 (p=0.02). In the prognostic cohort, no individual feature nor any SVM classification model was able to distinguish cases with a subsequent clinical bleed or biomarker event. Conclusion: This proof-of-concept suggests that ML may assist in deriving features on QSM maps acquired at a single timepoint, which reflect prior hemorrhagic activity in CCM. Further investigation is planned in larger cohorts and in conjunction with clinical trials of bleeding in CCM.