INCISE project and purchase of two high specification GPU cards
Computational workstations to enable deep learning-based bowel cancer screening in the University of Glasgow Digital Pathology Research Centre
We are delighted with the support from Cancer Research and Genetics UK which will facilitate the creation of a Digital Pathology Research Centre (DPRC) in the Institute of Cancer Sciences. As technology becomes available to improve cancer screening and diagnosis with the help of computers, we are building this centre to address this need in Glasgow and beyond. We will build on-site and remotely accessible digital workstations to further our research into cancer detection and classification with the purchase of two high specification graphics processing units (GPU) cards, connected by an interlink so they can be used together (giving a colossal 96Gb of GPU memory) or as separate cards (allowing two different scientists from two different groups to use them simultaneously). These computers are central to the deep learning aspirations of the DPRC and our first project around intestinal polyps and colorectal cancer, INCISE, which is described below.
Deep learning is used to infer conclusions from images. For example, it can examine pathology specimens and identify which cells are cancerous and whether they are actively spreading from the tumour. They can do some tasks far more accurately than a human pathologist and have the further advantage of not getting tired or bored, but we have found they are most useful when used as tools to help pathologists identify key tumour features. To do this, deep learning networks must be trained using huge amounts of data, with the amount they can hold at once determines their eventual accuracy.
Graphics processor computing is a highly specialised subfield of programming that has been embraced by the deep learning community; it is particularly suitable because the calculations that are needed to train models can be performed in parallel, making GPUs much faster than traditional computers. The computer we have requested will more than double our capacity for training new networks, which will make it possible to use more detailed and informative models. It will also be powerful enough that the computer would have been in the world’s top 200 or so as recently as 2012.
We will set up an on-site (located in the Wolfson Wohl Cancer Research Centre, WWCRC) and remotely accessible well-equipped computer with two NVIDIA RTX A6000 GPU cards, linked together using an NVLINK to create a machine in which both A6000s can be used together if large, data-hungry models are needed, or separately by two groups if not.
Our exemplar project for the DPRC is the innovative INCISE (INtegrated teChnologies for Improved polyp SurveillancE) programme, which aims to transform bowel cancer screening in the UK by developing a tool that can predict which patients with precancerous growths (polyps) in their bowels, will develop further polyps in the future. Professor Joanne Edwards, Professor of Translational Cancer Pathology, is leading the collaboration consisting of academics from across the College of Medical, Veterinary and Life Sciences, NHS colleagues and Scottish technology companies. INCISE will combine polyp tissue from colorectal cancer patients with data from the NHS Greater Glasgow and Clyde Scottish Bowel Cancer Screening Programme to train algorithms to predict patients’ future risks. The team will combine information about specific changes in polyp structure, as seen under the microscope and analysed using deep learning, with new analysis of the genetic mutations that cause polyps to grow. This comprehensive risk stratification tool will, for the first time, predict polyp recurrence by utilising the latest developments in digital pathology, machine learning and next generation sequencing. INCISE will improve cancer detection, whilst directly reducing the number of people requiring repeated colonoscopy, a costly, invasive, and unpleasant procedure.