SEIDOR – HORUS: New advanced retinal optical coherence tomography image processing techniques for machine learning
SEIDOR
Sector: Health
Business Case
As ophthalmologic pathologies are a global health problem, the creation of Artificial Intelligence (AI) models can be a huge step forward in this field. The HORUS project focuses on the standardisation of OCT image domains of different origins using generative AI models. This is important work, as standardising the domains may help to create more robust and accurate future models. In addition, work is being done to remove noise from the OCT images to obtain a cleaner image to work with.
Objectives
Develop AI models that allow: OCT images to be standardised so that they all have a similar domain. Noise in OCT images to be reduced Development of an interface that integrates the various modules and models
Use case
For the standardisation of domains, there being images of 3 different brands, a model was generated to convert the images from one brand to another, i.e. a total of 6 models. These models were created by applying a CycleGAN that allows the style of one image to be converted to the style of another using a non-paired dataset. For the case of Speckle noise reduction, two approaches were performed, on the one hand a CycleGAN, and on the other hand a Pix2Pix, thanks to a set of paired images.
Infrastructure
On premise training and inference in SEIDOR’s DPC, with 3 high performance machines with GPU.
Technology
AI technologies that generate written or spoken language, images or videos (generative AI) Image recognition/processing Text mining
Data
Private and anonymised datasets provided by BioAraba of OCT images.
