Abstract WP250: Leveraging a foundation model for artificial intelligence-based follow-up infarct segmentation in acute ischemic stroke

Background: Follow-up infarct volume (FIV) is a proposed surrogate endpoint for proof-of-concept clinical studies in acute ischemic stroke. Manual annotation of infarction on diffusion-weighted magnetic resonance imaging (DWI) is labor-intensive, costly, subject to high intra- and inter-observer variability. To address this, we evaluated the performance of a foundation model-derived approach to automated FIV segmentation. Methods: We developed BrainomixSAM2, an artificial intelligence tool for automated FIV segmentation on DWI-b1000. BrainomixSAM2 is based on MedSAM2, an adaptation of Meta’s foundation model “Segment Anything Model 2” (SAM2) optimized for medical images. We performed a two-stage fine-tuning to allow automated segmentation without prompting, trained on 344 patients. We compared BrainomixSAM2 performance with ground-truth expert raters and the current state-of-the-art automated approach (DeepISLES). Examples are provided (Figure 1). We performed validation on two independent stroke imaging registries: Cohort #1: 61 patients with suspected anterior circulation large-vessel occlusion (West Virginia University, US) Cohort #2: 59 patients with mixed anterior and posterior circulation stroke (University Hospital Královské Vinohrady, Czechia) Performance metrics included volumetric concordance (intraclass correlation coefficient (ICC)), and segmentation accuracy (Dice Similarity Index (DICE)). Results: In cohort #1 (51% females, median age=66[IQR=25], median FIV=30.5ml[37.6ml]), the volumetric agreement (ICC) between ground-truth and BrainomixSAM2 was 0.98 [0.96-0.99]. The average DICE was 80.7% (SD=10.2%) (Figure 2). In cohort #2 (39% females, median age=66[16], median FIV=3.9ml[14.62ml], 24% posterior circulation strokes), BrainomixSAM2 performed similarly to DeepISLES: ICC = 0.98 [0.97-0.99] vs. 0.99 [0.97-0.99]; DICE= 71.9% [20.5] vs. 71.5% [20.4]; p=0.22 (Figure 3). Conclusions: BrainomixSAM2 demonstrates excellent volumetric agreement with ground-truth annotations across diverse stroke cohorts. Segmentation accuracy was comparable to the current benchmark, DeepISLES, supporting its feasibility and reliability for automated FIV assessment. Wider validation and further head-to-head evaluations are warranted.