Deep learning for clinical decision suppor

Lead Research Organisation: University of Edinburgh
Department Name: Sch of Engineering

Abstract

Magnetic resonance imagining (RMI) units render visualisation of internal structures of body components using electromagnetic field, which induce resonance on atoms. The induction initiates electric field gradient, which is captured by an RMI to generate images. The method is preferred to computed tomography as it does not involve X-rays, which exposures the patient to radiation. However, the method is costlier to perform and requires further skilled medical stuff to operate the units.

Back in 2014, German hospitals have performed 10.6 million RMI scans and The Bureau of Labour Statistics has projected that within next ten years the need for MRI technologists will increase by 12%, which is almost twice as much comparing to the average growth rate of all occupations in the US market. The aim of the project is to address the market's need for specialised skill set by building a clinical decision support system, which will speed up the diagnosis time and lower its costs.

Problem statement

With increased pressure on the healthcare system by growing population and aging society, there is a need for combining medicine with automated systems to ensure throughput of patients through the medical procedures based on RMI scanning.

Proposed solution

The modern society is surrounded by different applications of machine learning, which ranges from recommendations of products based on customer's purchase history to recognition of objects in given images. Further, in recent years, the focus of the research community has been around representation learning, which allows an algorithm to detect the meaning of objects just by passing unprocessed data.
The proposed solution for building the decision system will be though supplying images from RMI units with annotated regions of interest. Moreover, the system will be trained to combine the patient's background information and associate similar cases from the institutional database. It is important to mention that in order a system to be reliable, it should be made in probabilistic manner so a level of confidence is always reported to the operator.
Current state of the field

The decision systems are already enhanced into software for radiologists, but they suffer from the lack of ability to combine multiple images to produce a diagnosis. Moreover, electronic health records (EHRs) are not standardised and they are protected by the privacy policies, which limits the access for the research community. The project will require to investigate and create similarity measurements between cases on weakly and non-annotated data, which combine imagining and non-imagining sources.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R513209/1 01/10/2018 30/09/2023
2106646 Studentship EP/R513209/1 01/09/2018 28/02/2022 Grzegorz Andrzej Jacencow