Purpose:
To automatically detect irregular shaped abnormalities, such as large retinal hemorrhages in fundus photographs, which account for one of the main causes of false negatives (FNs) in existing computer-aided detection or diagnosis (CAD) systems. Elimination of those FNs is highly desirable in developing screening systems which can be translated into practice.
Methods:
20 images from DRIVE database were used in training and 1200 images from MESSIDOR database in testing for supervised classification. Splat samples were created by segmenting the entire field of view into non-overlapping regions of homogeneous color. Splat-based feature vector, which consists of mean colors from different channels and contrasts with respect to neighboring splats, was formulated to discriminate its characteristics. A kNN classifier assigned a soft label to each splat indicating the likelihood of it being a part of a lesion. For rarely seen large retinal hemorrhages, as they share common features with blood vessels, splat labels were obtained from manually segmented vasculatures for the classifier to learn the appearance of "blood" and extract the most relevant features. Vessel splats flagged by automated vessel segmentation were removed from the output hemorrhage-likeliness map and image level decision was made by upgrading it to a single hemorrhage index.
Results:
The detector could separate blood/non-blood regions of both regular and irregular shapes. The top 20 of 1200 testing images with the highest hemorrhage index were reviewed by a retinal specialist (MDA) in masked fashion, and 18/20 were found to displayed large extensive hemorrhages, retinal membranes, extensive exudates and atrophic laser scars.
Conclusions:
Splat-based feature classification is efficient in detecting the presence of irregular shaped abnormalities in fundus images. The methodology might also be useful in eyes that show scarring or pre-retinal fibrosis. The outcome can be integrated into a comprehensive screening system to boost its overall performance and reduce the number of FNs.
Keywords: image processing • imaging/image analysis: non-clinical • retina