No abstract

Retinal vein segmentation is a crucial task that helps in the early detection of health problems, making it an essential area of research. With recent advancements in artificial intelligence, we can now develop highly reliable and efficient models for this task. CNN has been the traditional choice for image analysis tasks. However, the emergence of visual transformers with their unique attention mechanism has proved to be a game-changer. However, visual transformers require a large amount of data and computational power, making them unsuitable for tasks with limited data and resources. To deal with this constraint, we adapted the attention module of visual transformers and integrated it into a CNN-based UNET network, achieving superior performance compared to other models. The model achieved a 0.89 recall, 0.98 AUC, 0.97 accuracy, and 0.97 sensitivity on various datasets, including HRF, Drive, LES-AV, CHASE-DB1, Aria-A, Aria-D, Aria-C, IOSTAR, STARE and DRGAHIS. Moreover, the model can recognize blood vessels accurately, regardless of camera type or the original image resolution, ensuring that it generalizes well. This breakthrough in retinal vein segmentation could improve the early diagnosis of several health conditions.

Deep learning (DL) techniques achieve high performance in the detection of illnesses in retina images, but the major- ity of models are trained with different databases for solving one specific task. Consequently, there are currently no solutions that can be used for the detection/segmentation of a variety of illnesses in the retina in a single model. This research uses Transfer Learning (TL) to take advantage of previous knowledge generated during model training of ill- ness detection to segment lesions with encoder-decoder Convolutional Neural Networks (CNN), where the encoders are classical models like VGG-16 and ResNet50 or variants with attention modules. This shows that it is possible to use a general methodology using a single fundus image database for the detection/segmentation of a variety of reti- nal diseases achieving state-of-the-art results. This model could be in practice more valuable since it can be trained with a more realistic database containing a broad spectrum of diseases to detect/segment illnesses without sacrific- ing performance. TL can help achieve fast convergence if the samples in the main task (Classification) and sub-tasks (Segmentation) are similar. If this requirement is not fulfilled, the parameters start from scratch.