Unraveling the complexity: deep learning for imbalanced retinal lesion detection and multi-disease identification

Deep learning (DL) has been widely used to detect abnormalities in retinal image. Typically, this task has been focused on a specific domain, such as diseases related to glaucoma or diabetic retinopathy, for example. In this study, we propose to identify lesions associated with both diseases using a single base model, Cascade R-CNN, avoiding the use of multiple DL models. The task is complicated by the need for annotations in datasets related to damages in another domain for which it was created. In addition, the size and shape of objects and bias toward predominant classes are evident. Several techniques characterize this work, including soft labeling for mask predictions, normalized Wasserstein distance for handling small objects, and experiments in image sampling during training with cross-entropy loss combined with Online Hard Negative Mining or asymmetric loss. For result refinement, cluster-weighted with Distance IoU improved final predictions. Based on mean average precision (mAP), a standard metric in object detection models, the reported results were 0.46, and all experiments were conducted on the public DDR dataset. A detailed error analysis by category was provided. In conclusion, the feasibility of using a single model was demonstrated, while the techniques employed helped to increase mAP-related metrics. Our research provides novel insights into the use of retinal photographs for the prediction of systemic biomarkers associated with multiple diseases.