Computer Vision News - August 2020

Runner-up for Best Paper Award 24 Best of MIDL 2020 Raghavendra Selvan is a postdoc in theDepartment of Computer Science at the University of Copenhagen. His work on tensor networks for medical image classification was a runner-up in the Best Paper award at MIDL 2020. He speaks to us about his research and where he would like to take it next. Raghav sees this work as a bridge between tensor networks , which are predominantly seen in quantum physics applications, and medical image analysis , which is mostly related to machine learning-based applications. Tensornetworksare linearmodelswhich are very good at operating and handling data in high-dimensional spaces. Images, on the other hand, are already high-dimensional. Once you lift these images into high-dimensional spaces, the space is so massive that you cannot do any meaningful computations there. Tensor networks give a reasonable strategy for approximations in these exponentially high-dimensional spaces. The work incorporates two classical image concepts into tensor networks so that they work well with medical image data – local orderlessness and looking at the images at multiple resolutions . He believes it is the first work to apply adapted tensor networks to medical image data in this way. The model is based on deep learning platforms , which puts it at somewhat of a disadvantage in terms of efficiency, but Raghav thinks if they focus on a more efficient implementation, they could reduce computation time by 20 per cent. He tells us that one key benefit of the model is that it requires very limited resources: “These days, we have GPUs that need humungous memory, but we show in the paper that tensor networks use only a fraction of the GPU memory. In settings where you don’t have expensive hardware and you can’t scale up because of resource constraints, a model like this could help. I’m interested in low resource machine learning . This is the first method that shows with less than 10 per cent on a single GPU you can do what convolutional neural networks do with four GPUs .” Currently, the paper looks at 2D classification tasks, but 3D extensions are coming next, and he would like to Tensor Networks for Medical Image Classification