Feature Selection

This video will demonstrate feature selection steps in GC Image.

Note

Demos 4, 5, and 6 should be completed in sequence.

The public dataset used in this demo can be downloaded here:

Supporting files to help in following along with this quick start demo can be downloaded here:

1 Feature Pruning

After we have generated a feature template, we will open it for review and editing.

1.1 Cleaning the Feature Template

The cumulative chromatogram and feature template file used this step is included in the Chocolates Demo Processing Files linked above (Cumulative+Feature.gci).

  1. From the Image software, select File > Open Image....
  2. Browse to the save location of the auto feature template. Open the .gcfm folder, then the Feature folder, and finally select the Cumulative+Feature.gci file within. Click Open.
  3. From the toolbar, select the Graphic cursor mode.
  4. Click and drag to select feature regions (the red outlines) in the column bleed of the chromatogram.
    • Holding Shift while clicking or clicking and dragging will allow selection of additional feature regions.
  5. Press the Delete key and click Yes on the Delete Graphic confirmation dialog.
  6. After the column bleed feature regions have been removed, open the Blob Table from View > Blob Table.
  7. Select the Areas tab, click on the table, and press Ctrl+A to select all areas from the table.
  8. From the Blob Table toolbar, select Add to Template.
  9. Open the Template Table from View > Template Table.
  10. From the Template Table toolbar, select Save Template. Browse to your desired save location and click Save
  11. Close the Image software.

1.2 Reloading the Feature Template

  1. From the Investigator/Analysis software, select File > Open Analysis.
  2. Browse to the save location of your previously saved analysis. Click Open.
  3. From the menu, select File > Reload All
  4. Select the Apply an existing feature template option and click Browse.
  5. Browse to your modified feature template and click Open.
  6. Click OK to reload the images.
  7. After the images have been loaded, view the results by stepping through the list of images on the left of the window.
  8. From the menu, select File > Save Analysis. Browse to your desired save location, provide a file name, and click Save.

2 Evaluate Features with PCA

We can visually evaluate the quality of our features with Principal Component Analysis. A premade analysis file that can be used in this step is included in the Chocolates Demo Processing Files linked above (Chocolate.gca/).

  1. From the menu, select File > Assign Image Class.
  2. Assign a class labels to all images.
    • Select all images of a class by using Shift+Click to select a range.
    • Click Assign Class Label > New Label, type a label, and click OK
    • The Chocolates data set should have 10 classes with 4 images in each (Biscuit, Blank, Cafe, MangoPassionFruit, MintLime, Noisettes, Orange, Praline, Raspberry, Vanille).
  3. Select the Attributes tab, Areas from the View drop-down menu, and Percent Response from the list on the left.
  4. Select the PCA Chart tab
  5. Click the gear icon in the upper right corner. Select Replace missing values with: Zero (0). Click OK.
  6. Click the Run PCA button and wait for the PCA computation to complete.

3 Filter Features

To demonstrate the effect of feature filtering, we will perform PCA on a subset of the classes.

3.1 PCA on Class Subset

  1. From the Images tab, select an image from the MintLime class, and select File > Remove Image.
  2. Repeat until you have removed all images from the following classes: MintLime, Orange, Praline, and Raspberry.
  3. Select the Attributes tab, Areas from the View drop-down menu, and Percent Response from the list on the left.
  4. Select the PCA Chart tab
  5. Click the Run PCA button and wait for the PCA computation to complete.

Notice the remaining classes are still entangled.

3.2 Perform Feature Filtering

  1. From the menu, select Analysis > Multi-Class Statistics.
  2. From the Filter drop-down at the bottom of the window, select F Value.
  3. Select > from the operator drop-down and type 10 in the text field.
  4. Click the Run PCA button and wait for the PCA computation to complete.

Notice the classes are cleanly separated after the feature filtering.