Purchase this article with an account.
J.W. Chan, V. Sivagnanavel, C.C. Klaver, B. Lau, V.N. Chong, R. Smith; An Inter-Institutional Comparative Study and Validation of Computer-Aided Drusen Quantification . Invest. Ophthalmol. Vis. Sci. 2003;44(13):3654.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: We previously developed software based on Adobe Photoshop (Adobe Systems Inc., San Jose, CA) to quantify the percentage of drusen in digitized fundus pictures using macular background leveling. In this study, we assessed the performance of this system at two independent institutions and validated the results with expert graders who quantified drusen area by visual estimation. Methods: Digital high-resolution images of ten patients with significant numbers of drusen were selected for the study. Two graders from the Digital Fundus Photo Reading Center of Columbia University and two graders from King’s College Hospital quantified the percentage of drusen in the central and middle Wisconsin subfields using the software system. Likewise, three expert graders also estimated the percentage of drusen in the digital images by visual estimation in 5% intervals. The differences between the mean software gradings from each institution, and between the computer and human gradings were compared. Results: In the comparison of computerized readings, the mean difference over all images was 9.3, SD 12.1% for the central subfield and 6.4, SD 8.5% for the middle subfield. After a resolution of disagreements in choosing a threshold for segmentation in two images, the mean difference was 4.6, SD 6.7% for the middle subfield. The means of the three expert visual gradings had standard deviations of 0 to 17.6% for the central subfield and 0 to 10% for the middle subfield in all images. The mean differences between the software method and the humans over all images were 8.1, SD 9.9% and 3.6, SD 4.3% for the central and middle subfields, respectively. Conclusions: Two institutions independently used the same software to achieve comparable drusen measurements. Validation by expert graders of the software measurements showed small mean differences and comparable standard deviations. Quantification in the smaller central subfield was more sensitive to inclusion or exclusion of any single lesion. Limitations included a tendency to include retinal pigment epithelial atrophy as drusen.
This PDF is available to Subscribers Only