AI to predict Oesophageal cancer outcomes from PET/CT
Research type
Research Study
Full title
Can detailed analysis of PET/CT images support artificial intelligence approaches to predict the overall survival for upper GI cancer patients?
IRAS ID
277971
Contact name
Nicholas Vennart
Contact email
Sponsor organisation
Newcastle Upon Tyne Hospitals
Duration of Study in the UK
2 years, 6 months, 1 days
Research summary
Research Summary
We have access to a database of patients who have undergone treatment for Oesophageal cancer. As part of their pathway, they receive a variety of medical imaging tests. Medical imaging is used to identify and characterise the extent of the patient's disease. In addition to visual information, the images can be manipulated using sophisticated computer analysis to identify "non-visual" features of the images such as statistical relationships between individual pixels in the image. Furthermore, the emerging field of artificial intelligence is increasingly being used to assist clinicians in reviewing the images. We propose using artificial intelligence to identify which of these "non-visual" best predicts the patient treatment outcome. We eventually hope to use the computer program to view an image taken at diagnosis to predict how well treatment will work and ultimately provide better guidance on which treatment option will work best for that particular patient.
Summary of Results
We present a study investigating patients who have undergone treatment for cancer of the upper digestive system, e.g. oesophageal cancer. As part of the treatment pathway, patients receive a variety of medical imaging tests. Medical imaging is used to identify the extent of the patient’s disease. In addition to the visual information from medical images, sophisticated computer analysis can be used to look at the “non-visual” information in the image such as the relationships between individual pixels in the image. We set out to investigate 3 key questions:
1. Are any of these ‘non-visual’ image features can be linked with patient survival, without disease returning, for 2 years (locally considered a successful treatment)?
2. How are the ‘non-visual’ features affected by the way the images are acquired?
3. Can we use a computer program (artificial intelligence) to predict whether the patient will survive for 2 years after treatment, based on the non-visual features of the image acquired before treatment?
We analysed the images of 144 patients with oesophageal cancer and investigated 58 non-visual image features. We found that tumours which were more non-uniform on imaging generally did not survive as long as patients with more uniform tumours. The way the images are acquired can affect the non-visual features and so caution should be used when comparing results between patients and between other studies. Using artificial intelligence for this patient group showed some early promising results whereby we were able to predict, with 83% certainty, which patients would fail treatment within 2 years, based on the initial images however further work with much larger patient groups is required to confirm this.REC name
East of Scotland Research Ethics Service REC 1
REC reference
20/ES/0115
Date of REC Opinion
9 Nov 2020
REC opinion
Further Information Favourable Opinion