Drusen detection in OCT images with AMD using random forests

Abstract

The progress of age-related macular degeneration (AMD) can be evaluated through the quantification of drusen. In the literature, drusen are usually located by thresholding the distance between their limiting boundaries. However, Dufour et al. [1] proposed the detection of drusen by using textural features in conjunction with a random forest (RF). In this paper, we extend the work of Dufour et al. by adding new features and by using multi-label classification. The new features are the distance between the limiting boundaries of drusen and the respective wavelet coefficients. The area under the ROC curve (AUC) and Dice coefficient (DC) of our implementation of the method of Dufour et al. were, respectively, 92.03% and 69.72%, while our method obtained 92.81%, 71.88%. Moreover, the proposed method can identify individually drusen present in clusters. The individualized drusen information allows the exploration of new range of metrics that go beyond the total drusen area or volume, such as drusen size, which is used to categorize the stages of AMD. Additionally, we evaluated using automatic segmentations to compute the features. The automatic segmentations were applied in two settings: for test data alone and for both training and test data. The AUC and DC for first situation were, respectively, 92.42% and 71.31% and for the second case were 91.93% and 70.21%. These results suggest that the method is robust to variability of the segmentation of drusen's limiting boundaries.