Image Dataset Preprocessing Pipeline

The original dataset comprised 1778 Pentacam 4-maps of refractive display images. Three-hundred-and-four images were classified as clinical keratoconus (K), 584 images as early keratoconus (E), and 890 images as normal (N). All image preprocessing was performed using Python imaging library. All images were cropped, keeping only a square composite image showing the 4-maps without the color-scale bars. Then, images were scaled to 512 × 512 pixels and saved. To remove the background outside the 4-maps, the images were pasted over a gray 512 × 512 background with an intervening third parameter black mask image containing white circles overlapping the four circles of the 4-map display image, to present the 4-maps over a homogenous gray background. All images were then denoised to remove numeric and spatial landmark overlays, leaving only the color codes by iterating over black then white pixel values consecutively replacing the thresholded pixel values with the average value of the nearest neighboring pixels using a Python script. This obviated the use of conventional filters that produce a blurry image with loss of information. At this stage, we isolated a set of randomly chosen 90 images representing each class equally (30 images from each class) to be used as the test set for the classification DCNN. The remaining images (original training set) were used for class-wise training of the pix2pix cGAN and further training/validation of the classification DCNN. Figure 1 depicts image preprocessing steps.

Image preprocessing pipeline: A) anonymized raw image. B) cropped image. C) background cleanup. D) denoising black maps overlay. E) denoising white maps overlay.