GC Image Users' Guide

Image Processing Filters

Image processing filters are required to correct acquisition artifacts and to detect peaks. This chapter describes five image processing filters:

A preliminary step in GCxGC image processing is to optionally shift the phase of the image with respect to the start time of the second dimension. In GC Image, this is termed Shift Phase.
Removing the baseline (or background) level introduced by the system is an important step in precise quantification. In GC Image, this is termed Correct Baseline.
GC Image provides for point-wise Arithmetic Operations, which operate on a pixel-by-pixel basis.
It may be desirable to excise region(s) of the image (e.g., a region containing acquisition artifacts such as a solvent front) before analysis. In GC Image, this is termed Mask Pixels.
After correcting the baseline, the individual blobs of pixels associated with the separated chemicals can be detected and accurately quantified using the GC Image Detect Blobs operation.

The Shift Phase, Correct Baseline, Arithmetic Operations, Mask Pixels, and Detect Blobs operations are invoked through the Filter menu of the Image Viewer and via buttons on the Image Viewer tool bar. Two other operations available from the Filter menu of the Image Viewer are described in other chapters. Match Template, is described in the next chapter Analysis. Search Library, is described in chapter GCxGC-MS Data.

Shift Phase

A GCxGC image is created from a time-ordered stream of samples. Ideally, the maximum time required for each second column separation is less than the time period of the thermal modulation. Maintaining this condition across the whole of the GCxGC process requires proper temperature programming. Then, all of the samples of each second column separation are contained in a single column of the image.

Under these conditions, it is desirable to synchronize the time zero of the second column separations to the row zero of the image. However, depending on temporal relationship between the start of sampling and the thermal modulation period, post-acquisition phase shifting of the data may be required. Even if sampling and thermal modulation are synchronized, other variations can make shifting phase desirable.

As an example of the need for phase shifting, suppose that in one data set, sampling is initiated at the time of the start of the thermal modulation cycle (as in Figure 1.A below) and that in another data set, sampling is initiated at the time just before the start of the thermal modulation cycle (as in Figure 1.B below). If this is the only difference in data acquisition, then the peaks in the second image will be offset relative to the first image.

In this case, the phase could be aligned by padding one of the images so that the modulation period was consistent with respect to the columns of both images. Padding is required only in the first column at the beginning of the image. Then, the first column, which contains the padding, is excluded from analysis, as is the partial column shifted off the last column at the end of the image. An example of this phase shift operation is illustrated in Figure 1.C.

Figure 1.A: The start of sampling is synchronized with the start of the modulation cycle.

Figure 1.B: The start of sampling is not synchronized with the start of the modulation cycle.

Figure 1.C: The image columns from the unsynchronized image are brought back in alignment by padding the data.

In GC Image, the Shift Phase operation is invoked by either clicking the Shift Phase button on the Image Viewer tool bar or selecting the Shift Phase item from the Filter menu. The operation begins with a popup dialog box for entering the number of pixels to be shifted. The shift can be a positive or negative number for an upward or downward shift, respectively. The actual shift is set to the number of pixels specified by the user modulo the number of pixels in the second column separation. The interface also supports resetting to the original phase.

Figure 2.A illustrates an image before Shift Phase. The color scale is mapped over a narrow range to make the small blobs at the top and bottom of the image more visible. Note that some blobs wraparound to the bottom of the image. Figure 2.B shows the image after Shift Phase of -75 pixels. The blobs that were wrapped around to the bottom of the image are shifted back to the top of the image.

Figure 2.A: An image before Shift Phase has wraparound.

Figure 2.B: An image after Shift Phase corrects wraparound.

Correct Baseline

In gas chromatography, the signal peaks, which correspond to chemical constituents in the sample, rise above a baseline level in the output. Under controlled conditions, the baseline level consists primarily of the steady-state standing-current baseline in standard GC detectors and temperature-induced column-bleed which causes a rise in the signal in the later portions of temperature-programmed runs. Accurate quantification of the chemical-related peaks requires subtraction of the baseline level from the signal.

Figure 3 illustrates a perspective plot of an isolated peak rising to a maximum value of over 23 pico-amps. However, the baseline in that region of the image is more than 14 pico-amps, so the actual maximum peak height induced by the sample chemical is less than 10 pico-amps.

Figure 3: Perspective plot of a GCxGC sub-image containing an isolated blob peak.

In a simple model of the two-dimensional GC process, each image pixel produced by the system is the sum of:

A non-negative baseline offset value that is present even when there is no sample compound detected.
The signal due to the presence of the detected sample compound(s).
Random noise fluctuations (including quantization).

Under typical controlled conditions, the baseline offset values change relatively slowly over time, forming a slightly curving baseline across the image. The signal and noise fluctuate more rapidly over time and so can be separated from the slowly varying baseline offset.

The GC Image Correct Baseline operation estimates the baseline across the chromatographic image based on a few structural and statistical properties of the two-dimensional chromatographic process. Then, the baseline is subtracted from the image, producing a chromatograph in which the peaks rise above a zero-mean baseline level. (The noise is assumed to be zero-mean, in that any offset in the image is modeled in the baseline.)

To perform the Correct Baseline operation, either click the Correct Baseline button on the Image Viewer tool bar or select the Correct Baseline item from the Filter menu. The correction operation takes a brief time (typically no more than a few seconds), after which the current image is altered. After baseline correction, it may be desirable to re-colorize the image to reflect the new range of values. For details about this process, see "Background Removal and Peak Detection in Two-Dimensional Gas Chromatography", Reichenbach, Ni, Zhang, and Ledford, Journal of Chromatography A, 985(1-2):47-56, 2002.

Configure->Configure Settings on the Image Viewer menu bar provides four parametric values for Baseline Correction:

The algorithm divides the image into successive strides, with the assumption that in each stride there will be a deadband (a region where there are no sample eluents and the pixel values are determined by the baseline value and noise). One of the configurable parameters is the size of the stride. The stride can be set to 0.5 (meaning the stride is half the size of the second column dimension, under the assumption that there is a deadband in both halves of each second column separation) or to 1.0 (meaning the stride is equal to the second column dimension, under the assumption that there is a deadband in each second column separation). The default is 1.0. A larger stride size tends to produce an estimate of the baseline that is smaller valued.
The algorithm looks for a number of the smallest pixels in each stride, under the assumption that the smallest pixels are most probably in the deadband. The number of smallest pixels selected in each stride is one of the configurable parameters. The default number of deadband pixels expected in each stride is 5 and the minimum number is 3. A larger number of smallest pixels tends to produce an estimate of the baseline that is larger valued.
The algorithm performs some smoothing of its estimate of the baseline. The size of the filter window used for smoothing is a configurable parameter. The default size parameter for smoothing is 7 and the minimum size is 3. A larger size parameter for smoothing tends to produce an estimate of the baseline that is more slowly changing.
The algorithm estimates the expected range of baseline values plus noise using a configurable parameter of the ratio of the size of the range to the estimated standard deviation of the noise. The default ratio is a range that is 3.5 times the noise standard deviation. Note, after more testing the default may be increased. The minimum ratio is 1.0. A larger ratio tends to produce an estimate of the baseline that is larger valued.

The Filters menu also has an option to reset the image to its original baseline offset (Undo Baseline Correction).

Figure 4.A illustrates an image before baseline correction. The image is for a blank run (i.e., no sample) and the value range of the color map is set to be very small (13 picoamps to 15 picoamps) in order to highlight the small but clear increase in baseline value with time. Figure 4.B illustrates the same image after Correct Baseline (with a value range of -1 to 1). The rise in the baseline level has been removed. Note that baseline correction does not remove more quickly varying acquisition artifacts.

Figure 4.A: An image before Correct Baseline has a clear increase in baseline level from left to right.

Figure 4.B: An image after Correct Baseline corrects the baseline level.

Arithmetic Operations

GC Image provides for point-wise Arithmetic Operations, which operate on a pixel-by-pixel basis. The currently supported operations are addition, subtraction, and multiplication. The operands are the current image and either a scalar value applied to all pixels or a GC Image file specified by its filename. If a second image is specified, it must have the same size in pixels as the current image. The Arithmetic Operations popup is shown in Figure 5.

Figure 5: The Arithmetic Operations popup.

Point-wise subtraction of a so-called blank run (a chromatographic run with no sample input) can be used to remove background artifacts. Point-wise addition of images can be used to obtain an average chromatogram.

Mask Pixels

GC Image can set regions of pixels to a fixed value. This operation, called Mask Pixels, is useful for a variety of purposes including elimination of acquisition artifacts and selective analysis of data. For example, ASTM 2887 is a standard method for estimating the boiling range distribution of petroleum samples using gas chromatography. It may be desirable to eliminate regions of the image containing only column bleed and so more accurately determine the boiling range distribution of the chemicals in the sample. Similarly, it may be desirable to separately analyze alkanes and aromatic hydrocarbons.

The first step in Mask Pixels is to select a rectangular or polygonal region using the graphics tools (as described in the chapter Graphics). Then, to perform the Mask Pixels operation, either click the Mask Pixels button on the Image Viewer tool bar or select the Mask Pixels item from the Filter menu. The Mask Pixels presents a popup interface in which the user specifies the value of the pixels after masking and whether the pixels interior or exterior to the selected region are masked. Mask Pixels typically is performed after baseline correction and with masked pixels set to value 0. To remove an artifact, the typical operation is to outline the artifact with a polygon and then set the pixels inside the polygon to 0. To perform selective analysis, the typical operation is to outline the desired region with a polygon and then set the pixels outside the polygon to 0. The correction operation takes a brief time, after which the current image is altered. In this version, GC Image does not support undoing of Mask Pixels. The undo operation will be implemented in a later version.

Figure 6.A illustrates an image with a small horizontal stripe of bleed that has been outlined with a polygon. Figure 6.B illustrates the image after Mask Pixels, setting pixels in the interior of the polygon to 0. The bleed stripe is eliminated.

Figure 6.A: An image before Mask Pixels has a clear acquisition artifact.

Figure 6.B: An image after Mask Pixels eliminates the acquisition artifact.

Detect Blobs

GC Image can detect and quantify blob peaks in an image. To perform the Detect Blobs operation, either click the Detect Blobs button on the Image Viewer tool bar or select the Detect Blobs item from the Filter menu. The detection operation takes a brief time (typically no more than a few seconds) and does not change the image.

The Detect Blobs operation produces a table of blob attributes or features including peak location, area, volume, etc. This structure can be viewed and edited with the tools described in chapter Analysis.

The blob detection algorithm uses a greedy dilation that successively attaches the largest-valued unassigned pixel to a neighboring blob or forms its own blob if no neighboring peak has been established. Configure->Configure Settings on the Image Viewer menu bar provides parametric settings for Blob Detection:

Before the algorithm is run, the data values are smoothed to reduce spurious variations associated with noise. The smoothing is considered only during the detection algorithm and no pixel values are actually changed. The widths of the Gaussian blurring filter is configurable. The default width parameters are 0.5 pixels in the first column direction and 2.0 pixels in the second column dimension. The minimum values are 0.0. Larger values tend to produce more errors where two adjacent peaks are detected as a single peak, whereas smaller values tend to produce more errors where a single peak is detected as two separate peaks.
The blob detection algorithm implements a post-processing filter to eliminate detected blobs that have either an area (number of included pixels) that is too small or a maximum peak value that is too small. The default minimum area is 22 and the default minimum peak is 5.0 times the estimated standard deviation of the noise. The minimum value for these configurable parameters is 0.

Contents Previous: Graphics Next: Analysis