Automatic characterization and segmentation of classic choroidal neovascularization using Adaboost for supervised learning

Age-related macular degeneration (AMD) with the complication of choroidal neovascularization (CNV) has been the major cause of severe, irreversible vision loss in many developed countries. The imaging modality that is commonly used to reveal the vascular abnormalities of CNV is fluorescein angiography (FA). Analysis and interpretation of FA sequences are largely performed by skilled observers on single angiographic frames with considerable observer variability since much of the temporal information is lost. To facilitate accurate diagnosis of CNV, we propose a computer-aided diagnosis system which automatically characterizes regions of classic CNV in terms of severity by exploiting the characteristics of the temporal profile of fluorescence leakage, followed by automated delineation of CNV lesions. Our system is a supervised learning system which is trained with sequences of classic CNV to recognize the typical leakage characteristics of CNV lesions described using intensity values and intensity changes of a number of time intervals. The learning strategy is Adaboost, which effectively combines a number of weak classifiers to substantially improve the accuracy. Images of a FA sequence are first aligned and mapped to a global space, and the result of Adaboost classification is a severity map for the sequence - each point in the global space is assigned a degree of severity. CNV lesions are segmented using Random Walks algorithm, and the seed points are automatically generated using the severity map and labeled as either background or foreground based on the severity. We validated our algorithm using 16 classic CNV sequences. The best accuracy is 94% and the average is 75.5% with the major disagreement with the groundtruth coming from regions of fibrosis, which are treated as inactive CNV by our system but labeled as CNV by the retina specialist. If cases with high percentage of fibrosis are removed from the dataset, the average accuracy is 80.7%.