Resumen |
With the continuous increasing advances in hardware, there is a growing interest in the automation of clinical processes. In this sense, retinal blood vessels segmentation is a crucial step in the search of helping clinicians to get a better detection, diagnosis and treatment of many diseases. To solve this problem several solutions have been created, many of them using different deep learning architectures and with performances up to 95%. Some of these solutions need big datasets and they also use image preprocessing. In the present work we propose solving this problem with a cGAN on a small dataset and with color segmentation, making available distinguishing between arteries and veins, but more important we discuss how sometimes these high reported performances can be due to an improper use of the technic and this can lead to a not reliable model, bad reproducibility of results and non-sense comparatives with issues in the implementations and when used by clinicians. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG. |