Computer Vision News 30 Deep Learning for the Eyes further imaging is required, the closest facility that fulfills the requirements gets suggested. And the best part about the platform is that all the amazing deep learning research we all work on daily comes into action. Different algorithms are combined to analyze both image and text data to give additional support to clinicians. The result is a PDF-summary report with medical images, highlighting regions of interest alongside other clinical information. So what kind of algorithms are in use or currently in progress? The list is actually quite long and ranges from classification to segmentation up to the use of LLMs. Supervised, selfsupervised and semi-supervised learning strategies allow the use of labeled, unlabelled and weakly labeled images. Let’s start with classification. Grading of currently five main eye diseases in fundus images is (or will become soon) available in the tool: diabetic retinopathy, age-related macular degeneration, glaucoma, hypertension and cataracts. The approach used is quite simple: a ResNet-18 with smart data augmentation and with standard class activation maps. The main goal was to make it as robust as possible, specially to changes in the acquisition devices, which are expected to differ in the social context where the platform will be deployed. Hence, the team had to make sure to get as much variety as possible from different databases. Next, image quality assessment is of great importance for clinicians since “ungradable images” are … well … ungradable. Here, Nacho experienced that the current research doesn’t align with what clinicians want. Usually, research focuses on binary classification or in the best case classifying the images into usable, acceptable and rejectable. But technicians don’t like that. Instead, giving information on contrast, brightness and clarity of an image is better accepted, as it might help them to correct the acquisition protocol right away. Moving on to segmentation, the current focus is optic disc and cup segmentation, which plays an important role in the diagnosis of glaucoma. Nevertheless, the algorithms that Nacho and his team are creating will soon enough be extended to a variety of lesions, giving additional information on eye health.
RkJQdWJsaXNoZXIy NTc3NzU=