Convolutional neural networks (CNNs) show impressive performance for imageclassification and detection, extending heavily to the medical image domain.Nevertheless, medical experts are sceptical in these predictions as thenonlinear multilayer structure resulting in a classification outcome is notdirectly graspable. Recently, approaches have been shown which help the user tounderstand the discriminative regions within an image which are decisive forthe CNN to conclude to a certain class. Although these approaches could help tobuild trust in the CNNs predictions, they are only slightly shown to work withmedical image data which often poses a challenge as the decision for a classrelies on different lesion areas scattered around the entire image. Using theDiaretDB1 dataset, we show that on retina images different lesion areasfundamental for diabetic retinopathy are detected on an image level with highaccuracy, comparable or exceeding supervised methods. On lesion level, weachieve few false positives with high sensitivity, though, the network issolely trained on image-level labels which do not include information aboutexisting lesions. Classifying between diseased and healthy images, we achievean AUC of 0.954 on the DiaretDB1.
展开▼
