An Investigation into the analysis of Computed Tomography images using Deep Learning methods

Toshiba Medical Visualization Systems Europe, Ltd
Academic Supervisor: 
Dr. Paul Siebert

Deep Learning (DL) was hailed one of the ten breakthrough technologies of 2013. It extends the traditional neural network model by adding more layers, allowing higher degrees of abstraction and greatly improved prediction performance. In imaging applications, convolutional neural networks (CNNs) have brought about a step-change in the performance of many computer vision challenges (e.g. the ImageNet challenge).


A DL image analysis system can be trained on hundreds of thousands of cases, more than a human is likely to see in their entire career, and certainly more than they could draw upon in their decision making. However, off-the-shelf networks for other applications fail to exploit specific knowledge of the medical imaging domain.


Current machine learning approaches are almost completely data-driven. Without a doubt they work remarkably well in many situations. However, they are not the solution to everything. For example, if a physical system is well understood and can be described by a simple model, there is no need to waste effort mining for the model. Probably more pressing than time wasting though are situations where the data available to train a model is very sparse, and insufficient to train a purely data-driven model. While there are tricks to get around this to a certain extent, this project will seek to solve the problem by combining known physical models with data-driven training, greatly reducing the amount of data needed to train the models, and hopefully constraining it to be more physically feasible.


The key challenge in this project is to understand how best to apply DL methods to medical images in order to realise fully the potential benefits DL might afford to: (a) image reconstruction and (b) analysis and interpretation. Accordingly, we propose to investigate how to develop DL strategies to aid clinical diagnosis by taking into account the physical image formation process (including artifacts), noise properties, signal statistics and also clinical information that may be relevant, within DL-based analyses.