Classifying Diabetic Retinopathy Images Using Induced Deep Region of Interest Extraction

Classification and segmentation tasks on biomedical images have proven to be challenging and advocate the development of radically different approaches in machine learning. Machine learning methods involving deep neural networks have also acquired a phase-shift due to the inherent complexity associated with the representation, preprocessing, analysis, learning, inference and post-processing of bio-medical images. In this work, we consider the task of classifying diabetic retinopathy images coming from a small real-world data set. Taking advantage of the recent advances in bio-medical image segmentation to identify regions of interest, we propose a semi-automatic two-stage approach where the regions of interest identified by segmentation task fuel the classification task with additional discriminative information. We also devise a new loss function which improves the quality of segmentation tasks in diabetic retinopathy. Extensive experiments on segmentation demonstrate the utility of the proposed loss function. Further, classification experiments show a significant improvement in the generalization performance and validate the utility of the proposed two-stage approach.