Edinburgh Imaging

21 Mar 22. Deep learning in medical imaging lecture

Dr Maria Valdes Hernandez gave a lecture in Deep Learning in Medical Imaging, which is available to view on the Edinburgh Imaging site.

At the recent “Medical Image Analysis: application to brain” course, hosted by the University of Florence, Dr Maria Valdes Hernandez, Lecturer in Medical Imaging Analysis delivered her lecture: Deep learning in medical imaging.

 

The lecture begins by introducing the neuron as the basic computational unit of the brain. Dr Hernandez goes on to describe neuronal connections in neural pathways, or bundles, that are associated with different functions. She elaborates on how artificial neural networks attempt to mimic these natural neural pathways.

Following the introduction, the lecture presents mathematical insights of artificial neural networks and different types of network architectures. Dr Hernandez emphasises those schemes which have been most widely applied in the field of medical image analysis.

Some concepts within the computational jargon of artificial neural networks are also explained:

  • cost function or loss
  • activation function
  • gradient descend, and
  • backpropagation

The lecture presents and illustrates examples in MATLAB, of the pre-processing steps most commonly needed in brain MR image segmentation and classification tasks. These examples are extensible to other types of images, not just brain images.

The Open Neural Network Exchange (ONNX) format is described, which facilitates conversion between programming languages and platforms (for example, from MATLAB to python and vice versa). How to convert from ONNX format to the desired framework, or how to deploy the code for different platforms, using runtimes designed to accelerate inferencing, are also covered.

Dr Hernandez also explains the principles of convolutional neural networks (CNNs) and the elements involved in their design:

  • filter (kernel) size
  • number of layers and units
  • stride
  • pooling parameters
  • regularisation type, and
  • batch size

Generative Adversarial Networks (GANs) are also briefly introduced, as well as finally, different ways of validating the networks’ outputs.

The lecture concludes with some remarks on challenges that the field faces, and a list of platforms and academic journals where most of the advances in the field are published.

The lecture's content is directly applicable to the Edinburgh Imaging Academy’s Imaging MSc / Dip/  Cert, and Applied Medical Image Analysis Cert post-graduate programmes.

You can find out more information on the Edinburgh Imaging Academy degree programmes, here.

 

 

Social media tags and titles

Dr Maria Valdes Hernandez gave a lecture in Deep Learning in Medical Imaging, which is available to view on the Edinburgh Imaging site.

@EdinUniBrainSci @SVDResearch