GATE - Endoscopia Gastrointestinal Assistida por Computador Generalizável e Confiável

The GATE project aims to create the foundations of a new technology that can empower clinicians to better address the management of gastric cancer. Such technology, in order to be effective, requires algorithmic research beyond the state-of-the-art of today’s computer vision, which is only possible by bridging the gap between the areas of computer science and gastroenterology. In this section we will expand the first three project objectives listed in the previous section and identify the associated novel research challenges that will be explored in GATE. Since the fourth objective (early stage prototype) is more focused on experimental development, it will be detailed further in the description of Task 5 in section 8 of this proposal. Objective 1 - Manage and expand a multi-centric dataset of annotated gastric endoscopy exams, synchronised with other European initiatives on this topic. Real clinical scenarios have recently taught us that training supervised deep learning computer vision algorithms using large amounts of data from a single clinical setting (imaging equipment, local procedures, tendencies of the local medical staff) generalise very poorly to other settings. As such, multi-centric datasets are essential to obtain generalizable algorithms and in GATE we already have ethical and legal access to more than 30k real UGIE images from five different European clinical centres (IPO Porto, Rotterdam, Bucharest, Nantes and Barcelona), and a streamlined pipeline for its annotation based on a commercial specialised software. Furthermore, we have the support of two Data Collaborative European initiatives (EUCAIM, AIDA) that can help us maximise the international impact of these novel data annotations and algorithmic results. Objective 2 - Research and develop novel data augmentation algorithms focused on biologically viable image combinations that can generate very large synthetic annotated datasets. Besides using real data, we will take advantage of the multidisciplinary team of GATE to explore the creation of novel data augmentation algorithms that are able to produce images by combining geometrically transformed visual patches of lesions with normal UGIE images. The challenges here are how to choose biologically viable locations for the lesion patches, and how to visually blend them with the normal image. If successful, this will create an infinite generator of synthetic images with GIM lesions, which are already spatially annotated, boosting our ability to train deeper and more complex neural network architectures. Objective 3 - Research and develop novel deep learning computer vision algorithms, focused on generalisation and trustability, for the detection and quantification of intestinal metaplasia lesions. Understanding the average performance of algorithms is not enough to develop trust in an AI system. Reliability assessment approaches such as local fit and density principles try to address the problem of uncertainty in deep neural networks, enabling us to estimate the reliability of a single decision of this algorithm. Regarding generalisation, recent self-supervised learning paradigms have strong synergies with the multi-centric datasets of GATE, enabling the exploration of new approaches. These ideas will be expanded in the state-of-the-art section of this proposal.