accelerate spark data science residency, machine learning
How can we… use AI to advance metabolic medicine?
19 February 2024
Glioblastoma is a common brain cancer, yet there have been few significant advances in treatment for over a decade. For patients who are diagnosed with glioblastoma, the prognosis is poor. On average, they have just 15 to 20 months to live, even with an optimum combination of surgery, followed by chemotherapy and radiotherapy. During that time, their quality of life is negatively impacted. The challenge is to develop a new treatment that can improve patients’ prognosis in terms of survival and quality of life.
Symptoms of glioblastoma typically include headaches, cognitive problems and sometimes seizures. The cancer shows up on MRI scans as a big lump in the brain, but neuroscientists know it has usually infiltrated beyond what can be seen. When they operate, surgeons remove as much of the lump as is safely possible, while knowing it will probably not be the whole cancer. Currently, we do not have a lot of data to show how the surgery will affect patients ability to think and it is difficult to predict how long individual patients will live for, so we can’t give them any detailed information before the surgery. Despite chemotherapy and radiotherapy following surgery, in 90% of cases, the tumour will return within six months and despite our best efforts at treatment, patients deteriorate rapidly and then die.
Scanning for answers
In my lab in the Department of Clinical Neurosciences, we are trying to develop different types of MRI scans to better see tumours. This involves using AI to draw segmentations and calculate the volume of tumours, which can predict patient survival before their operation. Furthermore, if we can see tumours better, we can improve how to do surgery to remove more tumours safely.
We are also interested in using a battery of neuropsychology tests that link with imaging to measure cognition before and after surgery is carried out and build a model of how the brain and tumour interact. This method allows us to see which other parts of the brain may be affected by a tumour, giving us the option of performing surgery or giving radiotherapy to additional parts of the brain to improve the prognosis of patients.
Our current data shows that the majority of patients with glioblastoma already suffer from a degree of cognitive problems before their operation. In the first week after an operation, they will typically experience a worsening of that cognitive function, which may be related to swelling, medication or the fact they are in hospital. However, unexpectedly by four to six weeks after surgery when they start chemotherapy or radiotherapy, many patients improve compared to early after surgery. Despite this, they will never recover the cognition they had before surgery, so the damage the tumour has caused will not go away. This means that surgery may not be causing the permanent additional damage that we feared.
While this may sound pessimistic, having a better picture of the damage caused gives surgeons some cause for optimism, because we can model future treatment and possibly remove damaged parts of the brain that we know will not recover, which could enable us to treat the cancer more aggressively.
Another avenue to investigate is why some patients improve better than others. I hypothesise that it is related to the intimal damage that a tumour has done, but everyone is unique and behaves differently, which is why I am using mathematical techniques and AI to model the degree of cognitive improvement taking individual variability into account.
I come from a clinical background so machine learning, computational methods and coding were new to me before I started by PhD three years ago. So, the Accelerate Programme’s Python course, Data Science Residency and Machine Learning Academy were all invaluable. I wouldn’t have been able to analyse medical images in the way I do without learning how to code and how to implement the algorithms and computational techniques we discussed on the courses. It also taught me how to think about high-dimensional problems involving extracting meaning from different-sized datasets and variables such as decoding how the cognition with MRI.
Looking to the future
I am currently involved in setting up a clinical trial so we will be able to go beyond the standard care for glioblastoma, by using imaging biomarkers to show the invasion of a tumour. This will be the first time we have an imaging biomarker that could change how we perform surgery, which is exciting.
In the short to medium term, I want to combine my clinical career with academic research in brain tumours. I want to continue my research looking at developing biomarkers of tumour-brain interaction and tumour damage. Hopefully, by the time I’ve finished my neurosurgical clinical training, I can start planning a clinical trial to look at how we can do surgery or radiotherapy differently based on the effect of tumour damage on the brain.
AI is already being used to label tumours on CT scans and in radiotherapy planning for other cancers, so hopefully I can use the technology to contribute to significant advances in glioblastoma treatment that is badly needed.