Introduction
We worked with the NHS to explore how artificial intelligence
(AI) could help
radiologists quickly compare and assess computerised tomography
(CT) scans to enable
quicker diagnoses and improve patient outcomes.
The challenge
Radiologists at George Eliot Hospital NHS Trust, which serves
more than 300,000 people across
Warwickshire, Leicestershire and Coventry, perform
around 60 scans each day.
Many of these are related to cancer, and in most cases a
comparison with a previous scan is necessary to assess lesion
growth or shape changes. This manual alignment and comparison is
labour intensive, but no suitable automation tools exist.
Automating this process and improving alignment and overlay of
scans would enable slight changes in volume or new lesions to be
picked up more quickly. This includes a report guiding
radiologists to review or further evaluate particular regions.
Increased accuracy would improve patient safety and outcomes by
allowing faster diagnoses, and successful automation would also
save radiologists’ time.
The approach
Roke was the supplier selected from the Accelerated Capability
Environment’s (ACE) Vivace industry
and academia community for this commission and delivered a proof
of concept (POC) in just 12
weeks.
The POC uses AI to automatically overlay
images, compensating for factors including movement and breathing
as well as body composition changes frequently found in patients.
The tool calculates potential anomalies in three-dimensions,
allowing volume changes in lesions, or new lesions, to be quickly
identified.
It was trained on anonymised CT scan data from 100 patients
supplied by the hospital which also included machine data on body
part positioning and the angle of scan.
Roke developed the graphic user interface (GUI) visualisation tool by
breaking the problem down into 3 key areas covering alignment,
tissue sectioning, and anomaly detection, which Roke termed the
‘Holy Grail’.
The impact
As part of validation testing, lesions were successfully detected
in 7 out of 9 patients, and in 10 out of 17 images.
Roke also developed a masked data deep learning approach, trained
on human body composition, which fills in what it thinks should
make up any given area. This can then be compared to the actual
scan, with the difference between prediction and reality helping
identify a lesion and score the degree of anomaly.
