Recognising Diseases with Algorithms

As part of a project, biomedical engineering students looked into ways in which algorithms may be able to help radiologists to interrogate medical images.

Thanks to x-rays, ultrasonic and magnetic resonance imaging, doctors are able to recognise maladies within the body without operating. To do so, radiologists have to analyse scores of images – that takes time and concentration, although computers may be able to ease their workload in future. Jakob Dexl, Lisa-Marie Kirchner, Maximilian Reiser and Michael Uhl, all of whom are currently studying Biomedical Engineering at Landshut University of Applied Sciences, tested the extent to which specific algorithms are able to pre-sort MRI images of a skull. This would allow doctors to put more concentration into conspicuous images where the diagnoses was difficult. The Landshut-based Mühleninsel Radiology Practice provided MRI data from actual patients – completely anonymised, of course – for the project. The team also received support from Professor Dr. Andreas Lienemann as well as technical assistance from Cerner Deutschland GmbH, who supply IT services to the health sector.

Machine Learning: The Programme Learns on the Job

Modern machine learning algorithms, which train their diagnostic skills based on existing data in a similar way to radiologists, are used to classify images as either “ill” or “healthy”. As Professor Dr. Stefanie Remmele, the lecturer who supervised the team during their project and final assignments, explains: “one major challenge during this project was that even images from healthy patients differ strongly from one another, due, for example, to the presence of benign conditions or old age. So there’s a huge amount of heterogeneity involved. At the same time, radiologists sometime differentiate between ill and healthy based on the tiniest details, which makes it very difficult for the algorithms to classify the images correctly". Therefore, the students defined image excerpts and features on the basis of which the programme should differentiate between ill and healthy – for example, the shapes and sizes of brain structures. To achieve this, they fed the algorithm with images and data.

It already performed well in the first trial conducted in the course of the project. Yet, according to Dexl: “a lot of research into the subject will be required” before computer-based diagnostics will actually be capable of reducing the workload of radiologists. He and his team used just a single image per patient, for example – in reality, several images, such as different cross sections of the brain or different views, are captured and used for a diagnosis. Dexl’s colleague Reiser is certain that: “Research into machine learning will continue for many years, and the topic is coming into its own in medical engineering”.