| There are several novel applications of Deep Learning in Medical Imaging and especially in Ophthalmology in order to provide solutions to unmet clinical needs. The research presented in this paper concerns semantic segmentation of lesions regarding Diabetic Retinopathy. Most of the state-of-the-art papers nowadays use Convolutional Neural Networks, Fully Convolutional Networks, and UNETs, a modified version of Convolutional Neural Networks for segmentation tasks. The robustness of UNETs, in conjunction with transfer learning, has been the main strategy to tackle the limitations of the available public datasets. In this paper, the encoder of a UNET has been substituted by MobileNetV2, which constitutes a novel approach for segmenting Diabetic Retinopathy lesions. Results show that the proposed model, in hemorrhages and soft exudates lesions surpasses other similar attempts. In the proposed model, sensitivity reached 0.89 in hemorrhages and 0.97 in soft exudates. Another novelty of the paper is that the results are further analyzed at the lesion level, in contrast to the common pixel-level analysis met in the literature, something that favors a more intuitive evaluation of the model. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.