AI In Healthcare seminar: Dr Danilo Silva de Carvalho

Please join us for the latest in our AI in Healthcare seminars when Dr Danilo Silva de Carvalho will be discussing his work:

Seminar delivered by Dr Mauricio Jacobo Romero from CRUK

bio: I am a Principal Clinical Informatician (Researcher) at the Cancer Research UK National Biomarker Centre, at the University of Manchester, working on Safe and Explainable Artificial Intelligence (AI) architectures, with focus on Representation Learning. I hold a Ph.D. in Information Science from the Japan Advanced Institute of Science and Technology (JAIST), having worked as a systems analyst within Brazilian state oil company (Petrobras) on job safety analysis (JSA) and environmental licensing control systems, followed by several scientific and technological projects in the fields of Artificial Intelligence, Computational Linguistics and Information Systems at leading research institutions: UFRJ, FIOCRUZ (Brazil), Insight Centre (Ireland), JAIST (Japan) and the University of Manchester (UK).
Overview of the talk:
Experimental Cancer therapies, particularly targeted therapies, which are designed to affect cell signalling pathways can, in certain circumstances, have unintended off-target effects on the retina. For instance, ocular toxicities, such as serous retinal detachment and retinal vein occlusion, are observed in the treatment with several protein kinase inhibitors, such as MEK inhibitors. Such retinal abnormalities may be subtle and can go unnoticed without early detection. However, standard OCT interpretation by ophthalmologists is slow, costly, and not scalable across oncology settings. This increases the risk of permanent retinal damage with high impact to the patients’ quality of life. While research on automated OCT interpretation via Artificial Intelligence (AI) methods has produced systems with accurate pathology identification capabilities, the adoption of such technology by clinicians remains limited by their (in)ability to provide factual explanations justifying their outputs. In the context of the A-EYE trial study, we developed a method for automated OCT analysis (OCT-VAE) centred on the interpretability of its results, without substantial sacrifice of accuracy. This is achieved through training a Super Resolution CNN - VAE model for representation of OCT scan images on a low-dimensional (32:1 compression ratio) space in which the images can be: ① Accurately reconstructed; ② Coherently grouped by pathology (biomarkers); ③ Compared morphologically for similarity. In the talk I will briefly discuss how each one of those 3 objectives was accomplished.

The talk will be held in person, in the Emmeline Suite at the Christabel Pankhurst Institute at the University of Manchester, and also via a Teams link (sent out with the meeting details):

Know what’s Happening Next — before everyone else does.

Get App

Host or Publisher Jon Parkinson