Please read the following summary for a detailed explanation of what the AI project involves.
Artificial Intelligence (AI) for the classification and predicting outcomes of sarcoma:
helping pathologists to make better diagnoses faster.
Tumours are broken down into two different groups: benign or malignant.
While benign tumours can grow to a large size, they do not spread to other parts of the
body. This means that if they are completely removed, the patient can be cured of their
disease. Malignant tumours are also known as ‘cancer’. They have the potential to spread
to other parts of the body, although this does not always happen, but generally requires
further treatment. Thanks to dedicated research, new treatments are being developed
constantly, and survival continues to improve for many patients with cancer.
Cancers are classified into three major types
*About 85% of all cancers are carcinomas. The most common types arise from the lining
cells of organs such as the lung, gut, mouth, uterus, bladder, prostate, and breast.
*About 15% of all cancers are those of blood (leukaemia) and lymphoid tissue (lymphoma).
*Sarcomas are very rare and represent about 1% of all cancers. Sarcoma can present
anywhere in the body and therefore the pathologist who first sees the sample of the tumour
may not be a specialist in sarcoma pathology. There are also more than 70 different types of
sarcomas making the diagnosis even more challenging for many pathologists.
Advances in medicine and pathology
With rapid advances in medicine over the last two decades, the sheer volume of new
information means it can be difficult for pathologists to remember everything about all the
different malignant and benign tumour types. This can then delay the diagnosis being made
and can also lead to unnecessary and expensive ‘specialist tests’ being done which results in
wasted time and money. There is also a shortage of pathologists in the UK, and the
increasing workload could also lead to delays in diagnosis and treatment of cancer.
Tumour samples seen down a microscope are not always easy to tell apart.
It is not always straightforward for a pathologist to distinguish different tumour types. For
example, a melanoma (skin cancer) can look very similar to an ‘undifferentiated sarcoma’.
In some cases, it can also be difficult to distinguish a sarcoma from a benign tumour. For
example, a ‘nodular fasciitis’, which is referred to as a ‘sarcoma mimic’ can look like a
spindle cell sarcoma. However, the treatment of these two tumours is very different.
Making the correct diagnosis is vital to guide the surgeons and oncologists to undertake the
most suitable treatment of a specific tumour type for each patient. This is known as
‘personalised medical care’.
Artificial Intelligence
Artificial intelligence can be used to teach a computer how to recognise different types of
tumours. This technology will be used to help pathologists reach a diagnosis more quickly
and efficiently. In our AI research we plan to train the computer to distinguish about 50
different types of sarcomas and their mimics. The aim is for the computer to analyse the
images of a tumour and suggest a list of the most likely diagnoses which prompts the
pathologist to undertake only the relevant ‘specialist’ tests to confirm the diagnosis.
This would significantly reduce the workload of pathologists and laboratory staff. This would
give pathologists more time to think about the ‘difficult’ diagnoses, as well as to train the
next generation of pathologists. AI would also provide pathologists to undertake more
research with the aim of improving patient care and survival.
Chordoma and AI
We plan to include chordomas in our sarcoma AI research. Chordomas can look like other
tumours, such as chondrosarcoma, another bone tumour, and also like different types of
cancers, particularly cancer of the kidney. The long-term aim is to collect the images of
chordoma from around the UK from the last 40 years or so and to build a library which could
be used to train a computer model to help pathologists make the diagnosis of chordoma
correctly but also to identify if there are subgroups of chordoma that could benefit from
different treatments.
What is required for training a computer to suggest safe diagnoses?
For AI to be successful it requires thousands of slides to train the computer. Experience in
the field of AI has demonstrated that the larger the dataset, the more accurate the model.
What will we do with the UKRI funding for ‘Classifying Sarcomas’?
With the UKRI and other funds awarded we will collect slides from about 50,000 patients
with different sarcoma types including chordoma. Because sarcomas are so rare it is
necessary to collect the slides from all over the UK and beyond from the last few decades.
We will work with AI/ Computer Scientists to train the computer to recognise the different
types of sarcoma types and their mimics. For a pathologist to provide a correct diagnosis it is
often essential to know information about a patient and the tumour. For example, some
tumours only occur in children and nearly all chordomas are present in the spine - having
this type of information is important when training the computer.
Privacy and Trust in AI models for sarcoma
To ensure that details of patients and their families remain private but at the same time to
allow research to be undertaken, personal details of patients will be ‘hidden’/ anonymised
to researchers. For AI to be used in clinical practice, both patients and doctors must trust
the technology. This trust can be gained first by the rigorous training and testing of these AI
models prior to use in ‘the real world’, that is, being used in clinical practice. Furthermore, it
is important that the pathologist will ultimately be in the ‘driver’s seat’ and the final
diagnosis will rest on them. The pathologist will not be replaced by AI, instead they will use
it as a tool to increase efficiency and diagnostic accuracy.
What is next?
To ensure that trust between the patient, the medical and the scientific community is
developed; we would like to know your views. There will be patient representation on the
Steering Committee of this project to ensure that we hear about patients’ views and their
concerns. We will listen and respond to them.
This information from patients, their partners and their children can be fed back through
the Chordoma UK community.
