Identifying features of prior hemorrhage in cerebral cavernous malformations on quantitative susceptibility maps: a machine learning pilot study

Abstract

Features of new bleeding on conventional imaging in cerebral cavernous malformations (CCMs) often disappear after several weeks, yet the risk of rebleeding persists long thereafter. Increases in mean lesional quantitative susceptibility mapping (QSM) ≥ 6% on MRI during 1 year of prospective surveillance have been associated with new symptomatic hemorrhage (SH) during that period. The authors hypothesized that QSM at a single time point reflects features of hemorrhage in the prior year or potential bleeding in the subsequent year. Twenty-eight features were extracted from 265 QSM acquisitions in 120 patients enrolled in a prospective trial readiness project, and machine learning methods examined associations with SH and biomarker bleed (QSM increase ≥ 6%) in prior and subsequent years. QSM features including sum variance, variance, and correlation had lower average values in lesions with SH in the prior year (p < 0.05, false discovery rate corrected). A support-vector machine classifier recurrently selected sum average, mean lesional QSM, sphericity, and margin sharpness features to distinguish biomarker bleeds in the prior year (area under the curve = 0.61, 95% CI 0.52-0.70; p = 0.02). No QSM features were associated with a subsequent bleed. These results provide proof of concept that machine learning may derive features of QSM reflecting prior hemorrhagic activity, meriting further investigation. Clinical trial registration no.: NCT03652181 (ClinicalTrials.gov).

Keywords: bleeding; cavernous malformations; diagnostic technique; magnetic resonance imaging; quantitative susceptibility; vascular disorders.

FIG. 1.

Graph showing ROC curves of three significant features (sum variance, variance, and correlation) identified in categorizing CCM lesions with SH in the prior year. SE = standard error.

FIG. 2.

CCM lesion QSM ROIs representative of features that 1) had a significant difference in means between those patients with and without SH, and 2) had an AUC significantly different from random guessing. Features include variance, sum variance, and correlation. High-value examples were selected from the top 75% of the value range represented in the cohort, whereas low-value examples were selected from the lower 25%. Corresponding conventional MR images of the same lesions are shown. T1w = T1-weighted; T2w = T2-weighted.

FIG. 3.

A: Diagnostic biomarker event SVM classifier performance. Graph showing the ROC curve of a composite classifier developed using SVM with linear stepwise feature selection tasked with distinguishing lesions above and below the 6% biomarker bleed threshold. B: Individual features of the diagnostic biomarker event SVM classifier. Graph showing ROC curves displaying single features used in multiple folds of the cross-validation method, identifying potential key features for further examination. Sphericity and mean lesional QSM were utilized in 5 of 5 folds, while sum average and margin sharpness were used in 2 of 5 folds.

FIG. 4.

CCM lesion QSM ROIs representative of features utilized in multiple folds of the SVM classification method in differentiating those with biomarker bleed from those without an event. This includes sphericity and mean utilized in 5 folds, and sum average and margin sharpness used in 2 folds. High-value examples were selected from the top 75% of the value range represented in the cohort, whereas low-value examples were selected from the lower 25%. Corresponding conventional MR images of the same lesions are shown.