Deeplearning and Biomedicine Copy

Google’s big breakthrough in medical data science

The evening started with Dr Lucas Fleurens – an intensive care specialist at Amsterdam UMC  – introducing exciting updates from the world of medical data science.

The first was a breakthrough by Google DeepMind’s AI program, AlphaFold.  It cracked one of the toughest problems in science by predicting the 3D shapes of proteins, the fundamental molecules of life. Understanding how these proteins fold up could help usher in a new era of scientific and medical progress.

Next was an invitation to sign up for Amsterdam Hacking Health, Hacking Health Amsterdam is a hackathon in which unusual teams work on extraordinary solutions for improving care and welfare in the Amsterdam area. The event takes place from 11 to 13 April at Broedplaats Contact. Finally, Dr Fleurens notified attendees that the May edition of the meetup would be a special ‘Dating for Doctors and Data Scientists’ event. Doctors and data scientists are invited to pitch ideas and projects on how data science can help in a clinical setting.

Deep Learning to help accelerate MRI reconstruction

The night’s first presentation was given by Kai Lønning, a researcher at the Spinoza Institute for Neuroimaging. He introduced a project which uses Recurrent Inference Machines (RIM) to help accelerate MRI reconstruction. This allows his team to leverage the power of deep learning without explicit domain knowledge. RIM is a deep-learning model designed as a general inverse problem solver. It uses an iterative process in order to estimate what the true signal of an MRI looks like from a corrupted image. This process can help reduce the time it takes to process an MRI of a patient and so reduce the time they need to spend lying in the scanner.

Until now, the main method of accelerating MRI reconstruction has been using an algorithm known as Compressed Sensing (CS). CS exploits the fact that MRIs are compressible. But a suitable compression must be chosen beforehand, whereas the RIM finds its own compression implicitly. This is an advantage of deep learning, as the job of hand engineering feature extractions can be assigned to the neural network itself, which generally leads to more useful extractions. Lønning also showed that you don’t need a lot of data to train the model. He explained that his models had achieved good results based on data from only three subjects.

Using a variety of examples, Lønning showed that the RIM can accurately and efficiently reconstruct sparsely sampled MR-images at varying acceleration factors, and thereby producing reconstructions of higher quality than those produced by CS. To test this, Lønning’s team also asked a neuroradiologist to do a quality assessment test of brain images, who rated the images produced through CS lower than those produced by a RIM. Another advantage of the RIM is that it has been shown to work on an MR-Linac machine.  This can simultaneously generate magnetic resonance images and deliver X-ray radiation beams — allowing radiotherapy to be adjusted in real time and delivered more accurately and effectively.

Bioinformatics for biomedicine

Sanne Abeln, assistant professor of bioinformatics at Vrije Universiteit Amsterdam, gave the next presentation. Bioinformatics is the type of data science that works with molecular-scale data such as protein and genome data. This field has seen a big revolution in the last decade with the development of genome sequencers, which can process a person’s entire DNA in around one day to a week. These machines can measure everything from mutations in DNA to tumour DNA and mRNA, the intermediary between DNA and protein. The last 10 or so years has also seen a huge amount of bioinformatics data produced, Abeln said, “but often it’s very difficult and tricky to handle and we often don’t know what we’re looking for.”

The problem with the data is that there’s so much of it: at least 20,000 minimum observations of genes per sample, a figure that can rise to six million. Disseminating the information you need from this data is very difficult to do through classical statistical machinery methods as they need huge amounts of power, Abeln said, likening it to finding a needle in a haystack. But researchers need to pinpoint observations that can be significant to understanding, developing and treating disease.

Abeln used two examples to show that by using large public data resources, data mining and data processing to interpret focused experiments you can narrow down and select pathways (or sets of genes) that are relevant to a bioinformatic research study. Both examples used this strategy in research studies into colorectal cancer, the second-leading cause of cancer death worldwide, which is mainly treated by chemotherapy.

The first example Abeln introduced was her team’s attempt to discover if genome variation across all a colorectal cancer patient’s cells could affect disease progression, such as how fast a tumour grew or most likely site of metastasis. The second example was a project with the aim to identify structural variants in tumour DNA that disrupted the biology of the tumour. Abeln showed that by using focused experiments in combination with public medical databases they had managed to find significant associated genes for the site of metastasis and tumour progression and identify structured variants in tumours that determine the clinical relevance of recurrent breakpoint genes.

About Amsterdam Medical Data Science

The Amsterdam Medical Data Science Group meetup takes place on the third Tuesday of every month in the Delta Room at the VU University Medical Center Amsterdam’s Intensive Care Unit. The next edition will take place on April 16.

The event aims to bridge the gap between health professionals and data scientists by bringing both together in an informal setting for presentations and pizza. Since it’s launch in August 2018 it has steadily grown in popularity with more and more people attending to hear the latest news and innovations in medical data and spark new collaborations.

For more reports from previous Medical Data plus Pizza Meet-ups, click here.

The Amsterdam Medical Data Science Group meetings are supported by The Right Data Right Now consortium, which includes Amsterdam UMC, OLVG, Vrije Universiteit, Pacmed, and the Amsterdam Economic Board.

To find out more visit: https://www.meetup.com/amsterdam-medical-data-science.

The Amsterdam Medical Data Science Group meetings are supported by The Right Data Right Now consortium, which includes Amsterdam UMC, OLVG, Vrije Universiteit, Pacmed, and the Amsterdam Economic Board.