Abstract
Abstract: :
Purpose: To visualize and study artefacts in digital color photographs compared to conventional 35 mm color diapositive film and to examine their implications for automated fundus image analysis. Methods: The study included two single-chip mosaic color cameras and a three-chip RGB CCD camera which were used both in color and in monochrome mode, as well as a single-chip monochrome camera and a digitized version of color film fundus photographs. A single healthy subject underwent dilated fundus photography of one eye using each of these systems. Visual examination was made by two observers using a color cathode ray screen. Automated analysis was made using a prototype computer algorithm for automated vessel tracking. Results: Qualitatively, the resolution the three-chip RGB camera was slightly lower than the digitized color slides, but no distortion of details was identified. Photographs from single-chip mosaic cameras, which utilize an undersampling technique in combination with a variety of post-capture interpolation algorithms, were identifiable by diffusely reduced spatial resolution and/or high-frequency periodicity artifacts that increased the roughness of vessel contours. This phenomenon resulted in a beaded apperance of vessels with a diameter less than 50 µm. Both artifacts disappeared when the single-chip cameras were run in monochrome mode. It was evident that a variety of proprietory interpolation techniques are used, of which no documentation was provided by the manufacturers. Interpolation artifacts were most prominent in the red and blue channels, which are more undersampled than the green channel. When the automated vessels tracking algorithm was applied to such single-chip color photographs significant difficulty was seen in the delineation of vessel contours when compared to other types of fundus photographs. Conclusion: The spatial resolution of digital color fundus photographs is limited by the sensor technology, specific artefacts being produced by color single-chip image plane sensors. These artifacts do not only contribute a potential change in resolution but specific pattern reproduction artefacts may create obstacles for automated pattern recognition systems. This involves not only the detection of trunk vessel contours but it may also result in the production of circumscribed intensity depressions that mimick microaneurysms when small vessels near the resolution threshold are being imaged. This is a critical methodological issue because retinal microaneurysms are only about 1-3 pixels in diameter.
Keywords: 430 imaging/image analysis: clinical • 364 computational modeling • 429 image processing