Purpose: 
CT perfusion (CTP) is commonly used during acute ischemic stroke (AIS) to quantify infarcted core and salvageable penumbra. Discrimination is achieved by software using various mathematical and hierarchal models that tag voxels below or above a specific threshold. Bayesian-based deconvolution is a probabilistic method providing more accurate estimation of hemodynamic parameters than singular value decomposition [1,2], but like all methods is dependent on source dataset. This study aims to determine optimal CTP thresholds for identifying core infarct and penumbra using Bayesian modeling with two different smoothing filters and optimized dataset sampling. 

Materials and Methods: 
36 patients with clinical AIS, baseline CTP, cerebrovascular imaging, follow-up diffusion-weighted imaging (DWI) performed 18-48 hours after CTP were included. Imaging were  independently reviewed by a neuroradiologist not involved in original interpretation or endovascular procedure. Subjects had complete recanalization (23 TICI 2b/3) or non-recanalization (13 TICI 0) to assess core and penumbra respectively. CTP hemodynamic maps were computed using Bayesian-based deconvolution method with two different smoothing filters and automatic segmentations of core and penumbra volumes were performed applying combinations of best literature derived thresholds (CBF, CBV, MTT, TTP, Delay) to CTP datasets with absolute, relative or differential relationships using Olea Sphere® software v3.0SP20. Results were compared with manually segmented DWI final infarct volumes using automated non-rigid co-registration and resampling to provide volumetric and spatial accuracy. Receiver operating characteristic (ROC) curves with DICE and Youden indexes, volume correlation and differential weighted ranking of sensitivity and specificity were generated to assess predictive performance. Additional outside datasets (21 patients) with MRI performed at time of CTP or following therapy were used for visual differentiation of closely clustered candidates.

Results: 
Statistical and visual analysis showed infarct core volume was best defined by relative CBF < 25% and differential TTP > 5 seconds for the smoother filter and by relative CBF < 35% and differential TTP > 5.5 seconds for the other one; while penumbra volume by differential TTP > 5 seconds for the smoother filter and by differential TTP > 5.5 seconds for the other one. 

Conclusions: 
Consistent with literature, relative CBF is best predictor of core and improved with second parameter, TTP [3,4]. Differential or relative thresholds outperformed absolute values [5]. Weighting specificity over sensitivity improved correlation with visual ground truth indicating statistical superiority does not equate with clinical superiority. Bayesian TTP outperformed traditional time insensitive indicators due to novel influence of Bayesian modeling on deconvolution methods. Results indicate threshold selection within CTP software is critically dependent upon dataset filtering, convolutional methodology and the choice of absolute, relative or differential values and must be considered during product analysis.

CT Perfusion Imaging Using Bayesian-Based Deconvolution Method: Determination of the Optimal Threshold to Identify Core Infarct and Penumbra in Acute Ischemic Stroke

[Virtual Conference Page](https://event.vconferenceonline.com/microsite/html/login.aspx?id=1551&rlp=1)

[ASNR Website] (https://www.asnr.org//annualmeeting)

[Educational Exhibits](https://www.morressier.com/event/asnr2020exhibit/5e5d0be000701ceba4a26941)