Abstract Number: Esoc2026a556 NCCT-Based Automatic Detection of Severe Ischemic Hypodensity in Posterior Brain Structures to Predict Poor Functional Outcome After Basilar Thrombectomy

Abstract Background and aims Whether severe ischemic hypodensity (SIH) detected on non-contrast CT (NCCT) modulates endovascular treatment (EVT) effect in patients with a basilar artery occlusion (BAO) has not been described. We propose an automated method to quantify SIH in BAO patients based solely on NCCT. Methods We performed a retrospective international multicenter cohort study of BAO patients treated with EVT. An open-source deep learning model was used to segment posterior brain structures. NCCT scans from patients without posterior circulation stroke (n=342) were used to fit region-based linear estimators, to define patient-specific hypodensity thresholds based on the observed attenuation in the caudate nucleus. A hypodensity score was computed as the weighted sum of regions with high relative hypodensity volume; cerebellar hemispheres scored 1, while brainstem, midbrain and thalami scored 2. A score=2 identified patients with SIH. Univariate ordinal regression evaluated the association of SIH with clinical outcomes (modified Rankin scale, mRS at 90 days). Results Inclusion resulted in 171 patients (72 years, IQR 58-80, 42.7% female). Patients with SIH (33/171, 19.3%) had worse baseline NIHSS (p=0.029), but had comparable age, last-known-well to imaging times and mTICI2b/3 rates (p0.05). An ordinal mRS shift towards worse outcomes (OR 3.29, 95%CI 1.48-7.28, p=0.003) was observed in patients with SIH. Only 27.3% of patients presenting SIH achieved good functional outcome (mRS 0-3), compared to 53.6% in patients without (p=0.007). Conclusions Complementary treatments to EVT may be needed in BAO patients presenting early signs of SIH. Automatic hypodensity analysis could be used to help select patients in future clinical trials. Conflict of interest PC: Nothing to disclose. ATS: Nothing to disclose. PK: Nothing to disclose. NY: Nothing to disclose. HVV: Nothing to disclose. ML: Rapid AI. GWA: Rapid AI. SD: Rapid AI. MR: Methinks AI. JH: Rapid AI. Figure 1 - belongs to Conclusions