TPP

Proposal for 2019

Contact: ankit.sharma@tpp-uk.com

Final brief: Bone Doctor

discussion

X-Ray Machine Learning

Musculoskeletal conditions affect more than 1.7 billion people worldwide. These conditions are the most common cause of severe, long-term pain and disability, causing hundreds of millions of emergency department visits annually around the world. Successful diagnosis of musculoskeletal conditions currently requires X-ray analysis by skilled radiologists. In many parts of the world, however, access to these specialists is very limited. New advances in machine learning and medical imaging can help to solve this problem, improving outcomes for people worldwide. The task is to develop an algorithm capable of diagnosing these conditions at an expert-level, working with a new image dataset, and competing with teams from around the world.

The licence for the open data set use is here https://stanfordmlgroup.github.io/competitions/mura/. I am not sure if you want to check you are happy with the licence first?

response

We do often have students from our department participate in machine learning competitions of this kind, but those are generally PhD-level specialists in machine learning, using relatively powerful compute clusters. The group design project course is for second year undergraduates, with more limited computing resource, meaning that they would not practically be able to achieve competitive results.

As an alternative, in recent years, we have had a number of teams using pre-trained networks to implement application demonstrators in areas related to the original machine learning research but not competing with professional research teams. Do you think something like that might be possible, either working with the Stanford competition dataset, or with another medical condition? Even if pre-trained networks were not available, students could perhaps use this dataset to train their own network (likely achieving poorer performance than the Stanford benchmark), as the basis for a demonstrator in this area.

Here’s an example of a previous project from a couple of years ago that used a pre-trained network Neural Guide

Previous discussion

Contact: Sara Dowrick (sara.dowrick@tpp-uk.com).

Suggestion:

As healthcare resources become increasingly stretched, there is an ongoing push to enable patients to monitor their own health and wellbeing. The widespread use of smartphones has seen a large rise in the number of patient-facing medical apps available to users but encouraging uptake and use of these apps is a key challenge.

The aim of this project is to create a patient-facing, medical Android app that will improve, or aid in monitoring, the health of the user. In order to make the app more appealing and promote uptake, the app should be presented in a visually appealing 'gaming' format - e.g. a user is awarded points based on how well they adhere to their recommended salt intake etc.

In order to provide most benefit, special attention should also be given to the particular problem, disease or illness that the app has been developed to help with. A list of suitable causes will be provided to the project participants before commencing and will ensure there is plenty of scope for creative thinking and novel implementations...

Response:

It would be nice to do something in the healthcare area. I think it would be necessary to refine this a bit before advertising, as there are already a lot of apps in this general area. As an example of a more specific application, several years ago we had a team do a very successful project aimed at field support for patient records and monitoring during Ebola outbreaks.