PDQ-RA: Precise Digital Quantification of Retinal Angiopathy

PDQ-A: Precise Digital Quantification of Retinal Angiopathy: Method A Part Two: Prototyping the Method Using Actual Data L. Van Warren MS CS, AE Warren Design Vision September 5, 1997

Abstract
This paper follows the prequel in developing the method of PDQ-A, the precise digital quantification of retinal angiopathy, via separative colorimetry. The principal focus is that of prototyping the method using real retinal data. PDQ-A appears to be both extremely specific, and also adjustable. The method appears to perform better using real data, than in the synthetic case previously described.

Introduction
The labor intensive problem of accurate and precisely quantifying the extent of retinal neopathy was discussed in the prequel, along with a proposed seven step method, called PDQ-A. The goal of this note is to demonstrate that method using real retinal data. To this end, photomicrographs of actual prepared and stained retinas are analyzed below. For the purposes of this paper, the images are shown reduced from their full working resolution. This topic of resolution will be discussed briefly.

Resolution
Digital images have an associated spatial resolution of width by height pixels and a colorimetric resolution of colors possible shades given by 2^24 where 24 is three channels of color, red, green and blue, and eight bits of intensity for each channel - 3 x 8 =24. The image in this paper was digitized at 3001 by 980 pixels by 24 bits in an approximate aspect ratio of 3:2 from a 35 mm slide. It was then cropped according to the initial step of the method to 2987 by 1948 by 24 bits. The image was then spatially downsampled in an approximate 2:1 spatial ratio to 1493 by 974 by 24 bits prior to further analysis. The images published in this note are further decimated to 374 by 244 by 8 bits using an adaptive diffusion dithering method and stored for transmission using GIF compression. GIF compression causes some slightly noticeable color artifact and the decimation produces a significant loss of detail. This does not affect PDQ in any way, since full 24 bit color resolution is maintained during image analysis.

Step One
Image 1, below, shows that actual retina:

Actual retina image (courtesy Dr. John Penn UAMS). Slightly visible is a red spot in the upper right hand corner. This spot will be a significant feature later. This image was cropped from a larger field as described above, using an Ektachrome slide color corrected for tungsten light. No color bars are used in this initial experiment since we are primarily interested in relative color differences rather than absolute color values. We assume that the dynamic range of Ektachrome is adequate to capture the color relationships we are attempting to characterize. In summary, the Crop operator remains as step 1. We have additionally Downsampled for convenience.

Step Two
The goal of Step Two in the prequel was to eliminate non retinal areas of the scan from consideration. This was done by coloring non retinal areas with a null color. The null color is chosen such that it can be easily counted and subtracted from the slide area. Black was chosen on the synthetic retina, but white will be chosen here. Black is too close in color value to regions of profound neopathy. In the diseased retina selected here, the darkest regions of most profound neopathy have an {R, G, B} triplet value of approximately {9, 1, 0} with variation. Since there are no darker regions that are not neopathic we may set this extreme to {0, 0, 0}. Thus black is not convenient as a null color as it was in the prequel. Lighter, but still pathogenic regions have an RGB value of {80, 7, 0} with variation. The exact selection of a cutoff is greatly facilitated by inspecting the vasculature of the image and correlating the rgb values with the corresponding histograms. In particular observing the red and green image histograms produces a fascinating result. The green channel appears to be diagnostic, with regards to neopathy, since there is a local minimum that delineates normal vascularization from neovascularization. The red channel is less remarkable but still useful. Consider the following histograms for the active area of this image:

The darkest feeder vessels that radiate from the specimen center have a typical value of {107, 21, 0} with variation. Thus, as hoped, neopathy appears to have a worst case color separation of at least:

red: (107 - 80)/255 = 27/255 =10.6% green: (21- 7)/255 = 14/255 = 5.5% Blue is not a player. The exact statistical analysis of these variations throughout the normal and diseased retina is an important issue, whose ramifications unfolded during the course of this image processing experiment. Further analysis of the statistics, though interesting, will have to wait. We now return to the PDQ method.

The "orange peeling" conformal mapping of an approximately spherical retina to a flat slide is manifested here as a "four leaf clover" shape. In the raw data, non retinal regions of the scan have RGB color triplets in the range from {231, 247, 225} to {246, 253 237} As in the prequel we impose the SelectBoundary operator. But the BlackenRegion operation of the prequel is instead WhitenRegion for the reasons described above. This sets all pixels in the selected range to pure white or {255, 255, 255}.

Active retinal field. This change is barely perceptible in the illustration photograph, but is an important aid to accurate counting and thus is retained as we tailor our process.

Step Three
Image three, shows the result of an equalization step applied to the previous image. The equalization is only applied to those regions of image that are designated as belonging to the retina itself. The white background area is not included in the equalization.

Equalization of Active Field As in the prequel this step takes all the colors in the active region and distributes them further apart in color space. The step three operators are SelectRegion and EqualizeSpecial. Again this results in a barely perceptible change in the illustration image, but an important differentiation in the working image. Refer to the prequel to see a more pronounced example.

Step Four
Image 4, shows a selection step that will be used to identify those pixels in the active Field.

Selection of the Active Field As in the prequel this is a crucial step in the quantification process. In this image there are 363789 pixels in the entire rectangle, but only 347385 pixels in the active field. We are not interested in diluting our results with pixels that are not in the active field. The step four operator is SelectNotWhite. This selection of the active field was used to produce the histograms presented above.

Step Five
This step, which was specified in the prequel as, "moves all pixels within some color neighborhood towards blue". The choice of color neighborhood was not specified, but now, with the histogram results from a real retina, we have a precise analytic method for specifying which pixels move, i.e. those that are in the neopathic state. However instead of moving these pixels towards the blue for counting, we will move them towards their logical destination in color space: black. All pixels whose RGB values are in the intervaled set {0 - 80, 0 - 21, 0 - 255} go to black. Note that we are setting our red threshold as tightly as possible, i.e. as far away from normal vascularization, but we are allowing our green threshold to be set as loosely as possible, i.e. as close as possible to normal vascularization. This will, hopefully take advantage of the pattern we see in the shape of the red and green histograms respectively and be uniquely selective for pathogenic pixels. We are now at a pivotal and fascinating point in the process which will be noted as follows. When we began, we had an image in which it was easy for a person but difficult for a computer to discern the regions of neovascularization. By moving the pixels to black, we are now in the bizarre predicament of having an image in which it is hard for a person to differentiate but easy for a computer to "see". Nevertheless, at this moment, when viewing on a high resolution display, the remarkable and tunable selectivity of the PDQ-A method is evident.

Image 5 illustrates another critical step - the tagging of disease state regions.

Blackened Pathological Pixels In step six it will become easy for both human and machine to distinguish the critical regions. Note that with the exception of a small central region that none of the false positive artifact anticipated in the prequel has appeared. The step five operators are, SelectPixelPathosRange, and BlackenSelectedPixel. Automated counting of black pixels at this point will immediately produce the answer we seek. For the sake of the visualizing the result we will follow the prequel.

Step Six
Image 6a depicts the gradual removal of healthy region pixels from the scene.

Disease state pixels with envelope. It is now easy for both the human and the computer to see the key regions. The extreme selectivity of PDQ is also seen in this image. We have accomplished our intended colorimetric goal, "to move diseased tissue and healthy tissue to opposite extremes in color space". It is also important to note that we could adjust the "gain" on what the cut line between normal and pathogenic tissue is, simply by adjusting the cut lines in the histogram. This is a key enabling observation of PDQ-A.

We continue towards the goal of reducing the image to a count. This is pictured in image 6b:

Note that as a side effect we have identified not only the neopathic regions, but also regions that appeared damaged by other mechanisms, not relevant to our pursuits here. We inspect the histogram and discover that:

22,945 pixels have value {red = 0} 27,302 pixels have value {green = 0} 329,445 pixels have value {blue = 0} There were: 347,385 pixels in the active field Our image operators worked such that only diseased regions were moved to black. The condition for pixel black is that all three entries in the RGB color triple must be zero. This is true for only 22, 945 pixels.

We now have only those pixels we care about, the 22, 945 that represent regions of neopathy. We report for this retina that 22, 945 out of 347,385 or 6.6% of the active field pixels were affected, a precise quantitative measure compared to any manual estimation method.

Conclusion
The proposed method works. The next step is to apply the prototype technique to healthy, rather than pathologic retinas, to establish a baseline. After that comparison of manual estimation results with the automated version is appropriate. Following that the method should be reduced to a sequence of image processing operations requiring little or no human intervention. After that the method should be applied to series of digitized films in a run side by side with traditional techniques. After the method is stable, the film scanning step should be eliminated and replaced by direct color image acquisition. After that direct imaging of the eye, using appropriate dye strategies could be pursued. The PDQ-A method may have unanticipated benefits in the imaging of various regions of retinal insult in the living subject.

Acknowledgments
See the prequel.

References & Related Links (examined 9/5/97)
See the prequel.

To send email to the authors click on the links below: lvwarren at wdv dot com