Grant: Project Grant

Breast cancer is not a single disease and there are many different sub-types. It is important to understand what kind of cancer an individual patient has before deciding on treatment. Some breast cancers are not very aggressive and are unlikely to recur after the primary tumour is removed; for women with this kind of cancer it could be possible to avoid toxic chemotherapy drugs without affecting survival. Some subtypes of cancer can be identified by staining tissue sections from surgery with a special marker, but more recently, there have been many advances in personalized medicine based on the genetic makeup of the cancer cells. There are several genetic tests available that can help to differentiate between high risk and low-risk cancer, but these are very expensive and are therefore only made available to certain groups of women. They are also not 100% accurate and many patients end up with an intermediate risk score which is not helpful. More recently, advances in deep learning have led to some very encouraging work in digital pathology. A group from Google have shown that a computer can detect cancer cells more accurately than a pathologist and another group has shown that it is possible to predict the results from an expansive genomic test using a simple tissue slide image. Other work has demonstrated that using information from images can improve the accuracy of genetic tests. We have a large database of image data, pathologist's reports and genetic tests carried out on patients enrolled in a clinical trial. We also know which patients went on to have a recurrence of cancer. We propose to train deep learning algorithms to use the lower cost, more readily available image data to predict whether a particular cancer will lead to recurrence. We will compare the accuracy of these algorithms with the genetic tests. Our aim is to develop a tool for identifying high and low-risk cancers so that an individual patient can be offered the best course of therapy for them.