Deep multiple instance learning for automatic detection of diabetic retinopathy in retinal images

As a weakly supervised learning technique, multiple instance learning (MIL) has shown an advantage over supervised learning methods for automatic detection of diabetic retinopathy (DR): only the image-level annotation is needed to achieve both detection of DR images and DR lesions, making more graded and de-identified retinal images available for learning. However, the performance of existing studies on this technique is limited by the use of handcrafted features. The authors propose a deep MIL method for DR detection, which jointly learns features and classifiers from data and achieves a significant improvement on detecting DR images and their inside lesions. Specifically, a pre-trained convolutional neural network is adapted to achieve the patch-level DR estimation, and then global aggregation is used to make the classification of DR images. Further, the authors propose an end-to-end multi-scale scheme to better deal with the irregular DR lesions. For detection of DR images, they achieve an area under the ROC curve of 0.925 on a subset of a Kaggle dataset, and 0.960 on Messidor. For detection of DR lesions, they achieve an F1-score of 0.924 with sensitivity 0.995 and precision 0.863 on DIARETDB1 using the connected component-level validation.