Purpose:
The presence of retinal hemorrhages caused by Malarial Retinopathy can be decisive in the differential diagnosis of pediatric encephalopathy. We have developed an automated method for detection of retinal hemorrhages typical for malarial retinopathy.
Methods:
We conducted a cross validation study on an image dataset obtained from 14 subjects diagnosed with malarial retinopathy. A machine learning method was developed in which the data available from 12 subjects contributed in training and the data from remaining 2 subjects was utilized as a test set, in rotation. For each image, the imaged fundus is split into number of splats (over-segmentation of watershed regions) by ‘Tobogganing’ method. The variety of features based on RGB and HSV color channels, difference of Gaussian image and the adaptive histogram equalization image, are determined for each splat. A linear k-Nearest Neighbor classifier is trained with the splat features, which determines the properties of hemorrhages distinguishing them from non-hemorrhage regions and identifies the splats belonging to hemorrhages in a test dataset.
Results:
Figure shows a fundus image and the corresponding retinal hemorrhage detection (hemorrhage probability) image. The experiment resulted into an average sensitivity per splat and per lesion, of 74.15% and 97.53% respectively, on a ground truth validation. The system specificity was reported as 98.95% and area under the receiver operating characteristics curve was determined as 0.99.
Conclusions:
The proposed method provides automated detection of retinal hemorrhages for the diagnosis of malarial retinopathy. The detection performance of the system matches well with the ground truth. This suggests the potential of our method in providing an automated diagnostic assistance for malarial retinopathy screening.
Keywords: imaging/image analysis: non-clinical • lesion study • detection