MIDL Vision 2020

Ahmed Fetit is a Research Associate at Imperial College London, under the supervision o f Daniel Rueckert. This work is about trying to better understand why convolutional neural networks are so effective in segmentation tasks. Practically: how does the black box make such good decisions? He speaks to us ahead of his poster session today. The usual explanation given for why convolutional neural networks are so effective in many tasks is something called the shape hypothesis . Low-level features are combined in increasingly complex hierarchies into medium- and high-level features until an object can be classified or detected. These features are referred to as shape features . Things like lines and edges are based on something more complicated like squares and triangles. So, if you are classifying a cat, for example, you can detect the paw of the cat and the ear. This work suggests that although shape is an important element, it is not the only one. Image texture has a role too. By training the convolutional neural networks on explicit representations of the texture of an image, you can calculate texture maps and train the networks on them. Ahmed and his co- authors discovered this achieves good performance compared to just training them on an image. This supports their hypothesis that image texture is important, not only froma deep learning perspective, but from a neural imaging perspective too. He also believes it is the first time that this work has been done using neuroimaging data. The work uses a traditional computer vision method called Local Binary Patterns . Ahmed explains: “It’s computationally simple and straightforward to carry out. You represent each pixel by a number and this number is calculated by taking into account the local neighborhood of the pixel. Each pixel is a representation of the local textural variation. You end up with these simple texture maps that are very easy to compute. The great thing about them is that they are explicitly Poster Presentation 26 Training deep segmentation networks on texture-encoded input: application to neuroimaging of the developing neonatal brain