Machine-Learning-Assisted Detection of Residual or Recurrent Filling in Endovascularly Treated Intracranial Aneurysms on TOF-MRA: A Pilot Study

Background and Purpose: Surveillance TOF-MRA is a non-invasive alternative to digital subtraction angiography (DSA) for detecting aneurysm residual/recurrence after endovascular embolization but is limited by coil/WEB device-related susceptibility artifact and inter-reader variability. We evaluated a supervised machine-learning classifier for detecting residuals and/or recurrence on surveillance TOF-MRA obtained 3–6 months after endovascular embolization, referenced against the Raymond–Roy Occlusion Classification (RROC) class assigned on the surveillance TOF-MRA itself by a single, non-blinded neuroradiologist who had access to the completion DSA and 24–48 h post-procedural MRA. Materials and Methods: In this single-center retrospective pilot study, 53 of 55 eligible adults with intrasaccular endovascular embolization (coiling or WEB device, 2020–2024) were analyzed after excluding 2 for suboptimal MRA quality (RROC I: n = 13, II: n = 18, III: n = 27); the final analytical cohort comprised 58 aneurysms from 53 unique patients. Radiomic features from a class-agnostic 3 mm peri-device region were extracted with PyRadiomics and evaluated using L1-regularized logistic regression, calibrated linear SVM, and random forest. Model evaluation used 100 repetitions of 5-fold patient-grouped nested cross-validation, with all aneurysms from the same patient kept within the same fold and preprocessing, feature selection, hyperparameter tuning, and model fitting confined to training data. A post hoc exploratory subgroup analysis excluded the morphologically intermediate Class II cases. Results: In the primary analysis of 58 aneurysms from 53 patients, L1-regularized logistic regression yielded an aggregated repeated out-of-fold AUC of 0.716 (95% patient-cluster bootstrap CI 0.566–0.856), sensitivity 0.800, specificity 0.462, negative predictive value 0.400, and balanced accuracy 0.631 at a 0.50 threshold. The mean AUC across the 100 individual cross-validation repetitions, which is a different estimate, was 0.601 (SD 0.088) for logistic regression, 0.631 (SD 0.056) for random forest and 0.555 (SD 0.097) for the calibrated linear SVM. In the post hoc exploratory Class I versus III analysis (40 aneurysms from 38 patients), L1 logistic regression yielded AUC 0.832 (95% CI 0.704–0.946), sensitivity 0.852, specificity 0.692, and balanced accuracy 0.772. The subgroup represents a different, more separated classification task and is not directly comparable with the full cohort. Of 45 outcome-positive aneurysms, 42% were persistent residual filling and 58% interval recurrence. Conclusions: This pilot analysis demonstrates the technical feasibility of the patient-grouped radiomics-based classification of post-treatment surveillance TOF-MRA. The performance estimates remain preliminary because of the small single-center cohort, limited number of negative cases, scanner/device heterogeneity, and absence of external validation. The exploratory Class I versus III results should not be interpreted as evidence of improved clinical performance and warrant external multicenter validation. The specificity and negative predictive value were low, the incremental value of radiomics beyond simpler device- or morphology-based approaches has not been established, and the model is not ready for clinical use.