Evaluation of feature sensitivity to training data inaccuracy in detection of retinal lesions

Abstract

Computer aided diagnostic and segmentation tools have become increasingly important in reducing the workload of medical experts performing diagnosis, monitoring and documentation of various eye diseases such as age-related macular degeneration (AMD), diabetic retinopathy (DR) and glaucoma. Supervised methods have been developed for the segmentation and detection of lesions, and the reported performance has been good. The supervised methods, however, need representative data to properly train the classifier. Inaccuracies in the ground truth may have a significant impact on the performance of a supervised method as the training data are not representative. In this study, a quantitative evaluation of the sensitivity of different image features, including colour, texture, edge and higher-level features, to inaccuracy in the ground truth on exudates is presented. A mean decrease of approx. 20% in sensitivity and 13% in specificity was observed when using the most inaccurate training data.