Validating the Automatic Independent Component Analysis of DSA

Materials and Methods

This retrospective study was approved by the institutional review board of Taipei Veterans General Hospital. This study enrolled 36 patients who had extracranial internal carotid stenosis of >70% according to the NASCET criteria and who consequently underwent stent placement. DSA acquisitions were performed on a biplane angiosuite (Axiom Artis dBA; Siemens, Erlangen, Germany). A power injector (Liebel-Flarsheim Angiomat; Illumena, San Diego, California) was used to administer the contrast bolus. The image size was 960 × 960 or 1440 × 1440 pixels, and the FOV was 22 cm. The acquisition rate was 6 frames per second, and the acquisition lasted 8–18 seconds. Each patient was imaged in the posteroanterior (PA) and lateral views before and after treatment. A total of 144 datasets were analyzed.

The scale-invariant feature transform flow technique² was applied to the dynamic x-ray projection images for reducing motion artifacts. After performing the registration process, we used the FastICA technique (https://cran.r-project.org/web/packages/fastICA/fastICA.pdf)³ to segment the DSA images into arterial, capillary, and venous vasculature images. The thresholding method of Otsu⁴ was applied to the output independent component images for generating binary mask images. The binary masks of these 3 vessel types were used to measure the actual TDCs.

Each TDC was fitted by a γ-variate function⁵ to reduce random noise and recirculation. The following 4 parameters were calculated from the fitted curve: 1) the area under the fitted TDC (area under curve), 2) the maximum enhancement, 3) TTP, and 4) the width between the 2 time points on the TDC when the density was half the maximum enhancement, full width at half maximum (FWHM).

Three other parameters were calculated by fitting the γ-variate curve to straight lines. Every set of 4 consecutive temporal points on the γ-variate function was fitted to a straight line described by a linear equation: C(t) = mt + b, where m was the slope, using a least-squares error technique. The largest and smallest m values were recorded to represent the wash-in and washout slopes, respectively. The bolus arrival time was the time at which the straight line with the largest slope intercepted the horizontal axis (eg, Bolus Arrival Time = −b / Wash-In Slope).

Fifteen ROIs were manually selected to validate the proposed automatic technique. The locations of these ROIs are indicated in Fig 1. The first 13 ROIs had areas of 3 × 3 pixels. In the PA view, we selected 7 ROIs: the ICA at 2 locations (I1 and I2), anterior and middle cerebral arteries in the PA view, transverse sinus in the PA view, and ipsilateral and contralateral internal jugular veins (ipsilateral internal jugular vein and contralateral internal jugular vein in the PA view, respectively). In the lateral view, we selected 6 ROIs: the cervical ICA (anterior cerebral artery in the PA view [IA]), the cavernous segment of the ICA in the lateral view (IB), anterior and cerebral arteries in the lateral view, parietal vein, and superior sagittal sinus. The other 2 manually drawn ROIs were used to represent the manually selected capillary phase (MCP) in the PA and lateral views. In the MCP ROIs, pixels with maximum enhancement higher than the average were excluded.

• Download figure
• Open in new tab
• Download powerpoint

Fig 1.

Color-coded DSA images in the PA (A) and lateral (B) views. The numbers on the right-hand side are TTP values. Fifteen ROIs are manually selected to validate the automatic technique.

For each DSA dataset, 3 TDCs were obtained using the automatic method, and they were designated group A. A total of 15 TDCs were measured from the manually selected ROIs, and they were designated group B. Seven hemodynamic parameters, namely maximum enhancement, bolus arrival time (BAT), TTP, wash-in slope, wash out slope, FWHM, and area under curve, were calculated from the TDCs. We used the Pearson product-moment correlation coefficient to compare TDCs and these parameters between groups A and B. The cerebral circulation time (CCT) was defined as the time difference between the TTPs of IB and the parietal vein.¹ We compared the CCT with the FWHM measured from the automatically segmented capillary mask and the MCP ROI.