PDQ-RA: Precise Digital Quantification of Retinal Angiopathy

PDQ-A: Precise Digital Quantification of Retinal Angiopathy: Method A Part One: Description of Method L. Van Warren MS CS, AE Warren Design Vision September 5, 1997

Abstract
This paper describes PDQ-A, the precise digital quantification of retinal angiopathy, via method A, separative colorimetry. Retinal angiopathy is a leading cause of blindness. Various agents are continuously being evaluated for rectifying this condition. Accurate measurement of the effectiveness of these drugs is dependent on being able to quantify precisely the degree to which the disease process has been attenuated. A seven step image processing sequence is described that claims increased accuracy of measurement is described. This seven step process also has the potential to increase the speed with which the measurements can be made when compared to existing manual "clock face" estimation methods via automated image processing. The essential elements of the technique colorimetrically label diseased and healthy tissue imagery and separate these labeled types as far as possible in an RGB color space. Simple thresholding and histogram techniques are then used to compute the fraction of diseased tissue to healthy tissue.

Introduction
It is estimated that 85% of blindness is in some way traceable to angiogenesis, the neovascularization or intrusion of pathological capillary blood vessel development in the retina of the eye. This occurs in premature infants, in diabetics (proliferative diabetic neopathy), and others. It also occurs in tumors.

The process of finding effective drugs and clinical treatment schedules for minimizing angiogenesis is an arduous process. At the minimum the process consists of:
1) treating eyes with a candidate drug.
2) photographing retinas in a disease or pre-disease state.
3) manually assessing pathological regions on the retina using 12 slice "clock face" estimation.
4) comparing these results with other laboratory trials.

This process, first explained to me by Dr. John Penn, Director of The Center for Eye Research at UAMS, is lengthy and labor intensive. Penn estimates that step 3 alone requires 25% of the resources for evaluating potential treatments. I proposed that computing the areal density of angiopathic regions was functionally identical to the cloud removal problem in remote sensing; A problem I had first encountered as technical director of the first satellite composite view of the earth, done with Tom Van Sant of the GeoSphere project in 1990. Without going into excessive detail, the cloud removal problem in remote sensing is the problem of filtering out unwanted haze, clouds and suspended water vapor that obscure ones view of the earth's surface.

In remote sensing the problem is that clouds and haze bear a context-dependent similarity to ice and other features that occur in the image. Consider a contrived example. A haze obscured forest might have the same color as a slightly vegetated tundra. One wants to keep the tundra color unchanged while modifying the saturation of the forest to reflect its actual value. There is no global image processing operator, no global convolution kernel, that can accomplish this task. This problem is exacerbated by the fact that NOAA AVHRR data consists solely of three channels of wide band infra-red "water sensitive" data, and a visible red, and finally a visible green channel. No blue data was available for disambiguating earth surface scenes - a circumstance that persists to this day with the AVHRR family of orbiting sensors.

It turned out that manual intervention - later nicknamed "The Human Umpire Principle" - was necessary during image processing. This principle was manually and laboriously applied on a scene by scene basis, region by region basis, a sever violation of our original hope. A hope to completely automate the process of reconstructing a no cloud earth under a deadline from the National Geographic Society. Although the project was completed, incredible manual effort spanning months of time and the labor of several technical and artistic staff members was required. Later, in August of 1997, I faced this problem again when doing a second earth image and was able to do a reconstruction without manual intervention due to a change in colorimetric representation. Rendering the earth and the eye possess interesting similarities besides the fact that both start with E, and both are nearly spherical...

Back to Dr. Penn's problem: Retinas are photographed at high spatial resolution - approximately 1K by 1K or 1 Megapixel, but in gray levels from 0 - representing black to 255 - representing white. Although excellent for revealing retinal detail, monochrome imaging turns out to be less than optimal for differentiating regions of retinal pathology. The consequence was that lighter tones of the generally darker diseased regions have the same gray level as darker tones of the generally darker normal regions. Consequently - as in the cloud problem - there was no way of distinguishing between the two on a global basis. I proposed the equivalent of "we need a blue channel", which resulted in the simple option of color photography of the retinas. With multi-channel retinal data, distinguishing - and accurately computing the neovascular density - is perhaps within closer reach.

For the sake of this preliminary report, a simulated retinal image has been created. Although spatially the trial image possesses only a vague resemblance to a real retina, I propose that it is "close enough" colorimetrically for the purpose of process prototyping. This retina has been synthesized from a family of colors that correspond loosely to those that are found in real pathological retinas. For process design this is perfectly adequate. With that understood we will proceed.

Step One
Image one, below, shows that synthetic retina:

Synthetic retina image. This image was cropped from a larger simulated field, corresponding with that which would be available from color microscopic methods. So implicit with the appearance of the first image is a Crop operator.

Step Two
Image two, below shows the selection of the actual active retina in the field. The shape here is arbitrary, but the "orange peeling" is a significant factor in photographing actual retinas and so must be included here.

Active retinal field. This step requires selection of the relevant portion of the retina, called here "active retina" which includes normal and pathological regions only. All other portions of the image are zeroed. The dark tones suggest disease state while the red, pink and tan tones denote healthy tissue. The step two operators are SelectRegion, BlackenRegion.

Step Three
Image three, shows the result of an equalization step applied to the previous image.

Equalization of Active Field This step takes all the colors in the image and distributes them further apart in color space. The step three operator is EqualizeSpecial.

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

Selection of the Active Field This is a crucial step in the quantification process. In this image there are 58464 pixels in the entire rectangle, but only 30850 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 SelectNotBlack.

Step Five
Image five, illustrates another critical step - the tagging of disease state regions.

Tagging Pathological Pixels In this image we have identified pixels who are in the color space family of those designated as pathological. We have set a color space envelope that "moves all pixels within some color neighborhood towards blue". These pixels represent a selected disease state. The choice of this spectral direction is completely arbitrary and will likely change. The capture of specific pathological regions in PDQ-A is completely dependent on the pigmentation and colorimetry of the infiltrated region. Note in our current example that an artifact has appeared at the border. This represents a considerable area of "false positive" in this example. It is unknown whether this artifact will be significant while processing real retinal image data. The step five operators are, "SelectPathosPixel", and RecolorSelectedPixel.

Step Six
Image six depicts the removal of healthy region pixels from the scene.

Disease state pixels with envelope. Now we are left with only those pixels that represent retinal regions in a diseased state. This is the most sophisticated step in the process. The step six operators are, SelectNormPixel and BlackenSelectedPixel. We are nearly done.

Step Seven
Image seven shows the same image with the border artifact manually removed. Again it is not known at this writing whether the envelope will appear with a real retinal image. One should assume it will.

Disease state pixels only . This image now contains 227 different colors that are very close together in color space. The operator applied here is: RemoveEnvelopePixelsWithExtent. We are almost ready to obtain the final result, the data reduction of a diseased retina with a count of diseased pixels as compared to undiseased pixels - the critical metric for clinical analysis we were seeking.

Step Eight
Image eight shows the final resulting image:

Disease state binary image. We now have only those pixels we care about, the 316 pixels that represent regions of angiopathy. We report for this retina that 316 out of 30850 active field pixels were affected, a precise quantitative measure compared to any manual estimation method.

Conclusion
The process above prototypes one that will hopefully be applied to real retinas in the near future. It is hoped that this will not only improve the precision of quantifying angiogenesis, but that this technique will lend itself to automated processing of large sample sizes, enabling more potentially curative agents to be tested in a shorter period of time. It is also hoped that this method can be made real time by linking directly to multispectral digital CCD cameras connected to real time image processing platforms. This would enable "walk-in" and non-invasive methods of disease assessment and clinical evaluation. The PDQ-A method developed in this document is colorimetric in nature. A second complementary method B, called PDQ-B is stereographic in nature, exploiting the surface characteristics of neovascularization using stereographic imaging of retinas. Quoting Rocío Salceda of the National University of México, "The vertebrate retina consists of two components: the retinal pigment epithelium (RPE) and the neural retina itself. The RPE that originates from the optic cup, is a monolayer of cells located between the photoreceptors of the neural retina and the choroidal capillaries. The neural retina consists of six distinct neural cell types organized in a perfectly layered structure." Neovascular regions occur above the extremal retinal surface making three dimensional detection via elevation thresholding a possibility. If successful PDQ-A will be more precise in assessing the results of clinical trials than methods currently in use.

Acknowledgments
Without the following people this work would not exist. Kerrey Roberto helped me to understand the laboratory procedures that gave rise to the retinal images. Bethany Warren provides a daily fresh look at eyes. Lynn Warren prompted our first book on strabismus and reviewed this draft. John Peterson of Adobe for providing the excellent software used in this article. Netscape provided the browser used to lay this paper up and publish it instantly. Pat Kane of Champaign, IL provided enough RAM to process large images. Tsutomo Ohshima of Caltech put seriousness in my eyes. Tom Van Sant of the Geosphere project made a movie that started in space and ended in the eye of his son suggesting the connection between remote sensing and retinal imaging. Leo Blume assisted in the software implementation that led to the creation of the 4km GeoSphere image. Dave Warren, DA Hammond, Russ Sandberg and the late great Joseph Stone of the Little Rock Medical Center got me going. God made the earth and the eyes therein, the crown jewels of creation.

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