Digital image analysis

(1) Film scanner

Nikon 35mm film scanner LS-1000 with software, interfaced to a 100MHz Pentium Compaq PC. The LS-1000 is equipped with a 2,592 pixel linear CCD image sensor. The scanning area is 24.3x36.5mm, yielding (at maximum resolution) a 2,592x3,888 pixels array, for an effective scanning resolution of 9.4 µm square pixel. This results in a pixel density of 106 pixels/mm or 2,700 dpi (on film surface). Output data is 8 bits/color for a dynamic range of 256 levels. The image of a representative area captured by a camera was scanned (i.e. digitized) and saved in uncompressed TIFF format.

(2) Calibration

Standard (AO stage micrometer 2mm div. into 0.01mm) was used to calibrate pixel size measurement at the object plane. A positive image on Kodakchrome ASA 400 film was scanned at maximum scanner resolution (2,700 dpi). The number of pixels between micrometer graticules in vertical and horizontal directions (of slide) was averaged using time-domain and frequency-domain techniques. Pixel size at object plane was calculated to measure 0.1 µm².

(3) Image processing

Image processing was performed with custom software written in PVwave Advantage 5.0 from Visual Numerics, a fourth generation language for data visualization and processing running on a UNIX RISC workstation (DECstation 5000/200, ULTRIX V4.3).

Preprocessing

The resulting image files were on the order of 2-5 Mbyte in size. For purposes of efficient data manipulation and visualization, image files were decimated prior to quantification. A third-order, length 5 Savitzky-Golay filter was used to reduce noise artifacts in the resultant data prior to decimation. The decimation factor was chosen automatically to fit data into a fixed 800x800 pixels viewing window on a high resolution workstation monitor. The decimation operation preserves the aspect ratio, thus retaining a square pixel geometry.

Image segmentation

Irregular region-of-interest (ROI) areas were selected by the user via a custom designed graphical-user-interface (GUI). Due to anatomical variability in the image data, a manual boundary selection procedure was chosen. The contours of the intima/lesion border were outlined by an experienced operator using a computer mouse. Total area of ROI was calculated using enumeration of all pixels (function polyfillv, standard library PVwave, version 4.2) (23) within the user specified boundary. The resulting pixel count was then scaled by decimation factor to give actual area size in units of micrometers. For purposes of lipid quantification, each pixel in ROI was subsequently classified as either lipid or background, based on its luminance L, where L is an integer in the range 0 < L < 255. A standard binarization technique was used, based on visual inspection of the image. For each analyzed image, a global threshold value T was selected using interactive user input. An image pixel was classified as background if brightness value L < T, and as lipid otherwise. Total lipid area in micrometers was obtained by enumeration of all pixels in ROI with brightness L > T, followed by scaling with decimation factor .

Miscellaneous measurements

Additional measurements such as intima thickness, either as the distance between a single pair of points or as an average of multiple point pairs were performed. Point selection was made by the operator, using a mouse.

return to Introduction page