Non-parametric acute ischemic stroke penumbra delineation from dynamic DSC-MRI data with convex source separation
Sudhanya Chatterjee1, Dattesh D Shanbhag1, Uday Patil1, Venkata Veerendranadh Chebrolu1, and Rakesh Mullick1

1GE Global Research, Bangalore, India


In acute ischemic stroke (AIS), stroke volume is determined by DWI and volume at risk is identified by thresholding deconvolved Tmax map (> 6s). Tmax map is itself influenced by quality of AIF, its location, laterality and deconvolution algorithm. This can potentially impact estimation of "volume at risk”. In this work, we describe a CAMNS based source separation method with DSC concentration data to identify perfusion patterns without explicit parametrization of PWI data. We demonstrate that "volume at risk" estimation derived with CWSE may overcome the variability associated with the current methods based on Tmax maps only.


In acute ischemic stroke (AIS), diffusion (DWI) - perfusion MRI (DSC-MRI) are used to identify core and hypo-perfused areas [1]. With DSC-MRI, "tissue at risk" is typically identified by thresholding Tmax map [1]. Computation of Tmax map can be influenced by quality of AIF, its location, laterality and the deconvolution algorithm used [2, 3]. Consequently, this variability in Tmax will impact estimation of “volume at risk”. An approach to assess perfusion in AIS without explicit parametrization of DSC data is desirable. Source separation methods can potentially identify patterns from dynamic data[4, 5, 6]. We describe a method to identify perfusion patterns from DSC-MRI using Convex Analysis of Mixtures of Nonnegative Sources (CAMNS) based source separation [7] and weight estimation approach (CSWE) [6, 8]. We demonstrate that, CSWE can estimate reproducible, non-negative and physiologically plausible perfusion patterns from DSC data. The work indicates that "volume at risk" estimation derived with CWSE may overcome the variability associated with the current methods based on Tmax maps only.


Patient database: MRI data from ten AIS patients acquired with IRB approval. Imaging: Performed on 3T (N =8) /1.5T (N=2) GE Signa HDx MRI scanners using 8-channel head coil. DWI imaging: Axial DWI trace images, SE-EPI, b = 0 s/mm2 and 1000 s/mm2. PWI Imaging: Axial oblique slices, GE-EPI, 240 mm FOV, 128x128 matrix and, 3T: TE/TR = 19/1000 ms, 100 bolus phases, TH = 7 mm, 1.5T: TE/TR = 60/2000 ms, 25 bolus phases, TH = 6 mm. PWI Map Generation: DSC images were processed using BrainStat-AIF tool (Advantage Workstation, GEHC) to generate deconvolved Tmax and MTT maps. DSC images were also processed using in-house software to convert DSC signal data into gadolinium concentration units. Brain mask was segmented on first non-saturated bolus phase. CSWE methodology: CAMNS algorithm identifies underlying non-negative sources of concentration data within brain mask. We hypothesized three components (p =3): normal brain, delayed, low-perfusion and noise/core infarct . Post CAMNS, a constrained optimization is solved to calculate weights corresponding to sources as [7]: If data is $$$ X \varepsilon \Re^ {m x n} $$$ , (m = bolus phases, n = no. of voxels) and p-sources $$$ s \varepsilon \Re^{m x p} $$$, weights w calculated by solving $$$ \min_{w} \left \| X-ws \right \| $$$ such that w $$$ \geqslant 0 $$$.

Hypoperfused area delineation: In first pass (FP-CSWE), normal brain region source component was manually identified and its spatial weight data thresholded (< 0.5) to identify voxels which underwent second pass of CSWE (SP-CSWE). SP-CSWE sources and weight matrix were correlated to identify regions which corresponded to elevated Tmax regions. The SP-CSWE – Tmax correlation was performed by a trained radiologist who reviewed DWI, Tmax, MTT and SP-CSWE weight map data and identified SP-CSWE weight map corresponding to Tmax map characteristics. Next, SP-CSWE weight map corresponding to Tmax map was thresholded (> 0.95), ventricle regions (if any) removed by the radiologist and a binary mask generated.

Results and Discussion

Figures 1-5 demonstrate the efficacy of proposed CSWE method in identifying abnormal perfusion regions in AIS. In all cases, radiologist could correlate at-least one weight matrix component with elevated Tmax regions. We also investigated if certain trends are visible in the shapes of sources. One common trend was that normal region source had least offset from baseline; post the peak and re-circulation cut-off, while in case of elevated Tmax regions, the source had a significant offset and/or no “washout” of contrast (Fig. 2). For patient #5 (Figure, 5), Tmax map showed elevated values only around regions of DWI lesion in right PCA distribution, while CSWE data indicated that whole of right hemisphere and left PCA distribution being compromised. Review of MTT map by radiologist (Figure 5) confirmed this was reasonable; demonstrating the robustness of CSWE methodology. The only false positive in the correlated SP-CSWE binary mask was contribution from ventricles, which had to be manually removed. However with availability of DWI data (inherent suppression of ventricles in DWI) in AIS, this can be easily resolved by registration of DSC-MRI to DWI. In all cases (except one, fig 5), the SP-CSWE binary mask had mean Tmax values > 6s, similar to that observed in literature for AIS lesions [1]. The follow-up to this work would be to correlate the patterns with the final infarct outcome to confirm the robustness.


In AIS setting, CSWE methodology with DSC concentration data can reliably identify regions of impaired perfusion similar to that observed using parametric Tmax map.


No acknowledgement found.


[1]. Wheeler HM , Stroke. 2013 Mar;44(3):681-5. [2]. Calamante F, Stroke 2002; 33: 1146–1151., [3]. Kudo K, J Cereb Blood Flow Metab. 2011 Mar;31(3):908-12 [4]. Suzuki K et al, J Neuroimaging. 2011, 21(4):384-94. [5]. Li Chen, IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 12, 2011 [6]. Wu Y, Journal of Cerebral Blood Flow & Metabolism (2007) 27, 632–645 [7]. Chan AH, IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 10, PART 2, PP. 5120-5134, OCT. 2008 [8]. Palomar, Daniel P., and Yonina C. Eldar. , Convex optimization in signal processing and communications. Cambridge university press, 2010


Figure 1. Patient #1

Figure 2. Patient #2

Figure 3. Patient #3

Figure 4. Patient #4

Figure 5. Patient #5

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)