UCL Centre for Medical Image Computing


Tom Whyntie - Brice Thurin

27 September 2017, 1:00 pm–2:00 pm

Event Information

Open to



UCL Bloomsbury - Drayton House B20 Jevons LT (junction of Euston Road and Gordon Street)

Brice Thurin

Title: Light field medical imaging: applications in ophthalmic diagnosis and surgery


Light in a scene is transported, occluded, and filtered by its complex interaction with objects. By the time, it reaches our eyes, its radiance is an intricate function that carries information on all those interactions. Light-field cameras capture the spatio-angular distribution of light and enable one to reconstruct the radiance function. This, in turn, can be used to digitally refocus at different depths, remove occlusions or specular reflections, estimate scene depth and provide 3D visualisation. All these capabilities are crucial to endoscopically guided surgery. In this talk, I will focus on the application of this technology to ophthalmic retinal imaging. I will present the proof of concept light-field optical system we are developing. Further, I hope to spur the interest of the audience by highlighting the processing challenges and opportunities this imaging modalities represents.

Tom Whyntie

Title: Anomaly Detection in MRI brain scans using GIF-derived features


Machine learning has been successfully applied to the automation of many tasks in medical imaging, such as image classification and segmentation.  We are investigating anomaly detection - spotting anything odd in a given subject - not with images, but with brain volume measurements obtained from thousands of clinical-grade MRI scans processed with the Geodesic Information Flow (GIF) technique.  One such approach is based on the Autoencoder - a symmetric, deep neural network that reconstructs an input based on a lower-dimensional encoding of the GIF volumes.  If trained appropriately with a large enough dataset, the Autoencoder should be able to reconstruct normal, healthy samples well, but fail when presented with pathological samples.  The degree of failure provides a metric for abnormality that could, in principle act as a trigger system for further clinical analysis.  I will present some preliminary results along with a discussion of the challenges faced when processing such large datasets and defining what is "normal" anyway.