Automatic classification of nuclear medicine scans using deep learning
Research type
Research Study
Full title
Automatic classification of nuclear medicine scans using deep learning
IRAS ID
287927
Contact name
Patrick Fielding
Contact email
Sponsor organisation
Cwm Taf Morgannwg University Health Board
Duration of Study in the UK
1 years, 11 months, 30 days
Research summary
Deep learning is a branch of artificial intelligence that involves the application of a computational neural network to learn and make predictions about a dataset. A neural network is a model that consists of a number of computational nodes arranged in many layers, through which data is repeatedly passed. With each iteration, the neural network makes adjustments to a set of parameters associated with each node in an effort to optimise the fit of the model to the data. By feeding the network with a large volume of pre-processed data, it is possible to train it to recognise patterns within the data, and subsequently to make predictions about unprocessed data. For example, a neural network can be trained using nuclear medicine images that have been interpreted and labelled by expert Consultant Radiologists. The neural network should then be able to prospectively make diagnostic predictions about radiological images independently of human interpretation.
The aim here is to construct and train a neural network to classify nuclear medicine images according to a set of diagnostic codes that are utilised by Consultant Radiologists specialising in Nuclear Medicine in South Wales. All nuclear medicine studies reported by Consultant Radiologists contain an alphanumeric code in the report according to the interpretation of the study. There is several years’ worth of data available across Health Boards in Wales, consisting of hundreds of cases, which can be accessed via the Patient Archive and Communication System (PACS) system. Neural networks require a large amount of data to train. A proportion of this pre-processed data will be used to train a neural network to classify scans according to the appropriate diagnostic code. The performance of the neural network will then be tested against the gold standard Consultant Radiologist report using the remainder of the data.
REC name
East Midlands - Nottingham 2 Research Ethics Committee
REC reference
20/EM/0259
Date of REC Opinion
20 Oct 2020
REC opinion
Favourable Opinion