From brain surgery to AI for medicine

26 May 2023

Mastering brain surgery would be enough of an achievement for most people, but Dr Chao Li set his sights on using AI to improve the management of patients diagnosed with brain diseases. Dr Li is a Principal Research Fellow at the Department of Applied Mathematics and Theoretical Physics, and Department of Clinical Neurosciences, Group Leader of Intelligent Neuroimaging for Computational Neuroscience and Oncology at the University of Cambridge, and teaches fellow researchers how to harness the remarkable powers and potential of AI to understand the brain and solve clinical problems.

Beginning with the brain

Dr Li started his medical training more than 15 years ago, having become interested in the brain and understanding how it is affected by diseases. He mastered skills while working as a neurosurgeon at major medical centres in Shanghai and over approximately 10 years, performed thousands of operations to treat different neurosurgical conditions. “Brain surgery usually takes a long time, and in some cases, it requires fast reactions to save lives,” he said.

While he is passionate about making a difference to patients with his clinical practice, he also wanted to know more about how best to improve the standard of care and benefit more patients. He was particularly keen to understand more about malignant brain tumours. Currently, patients with a malignant brain tumour often have a poor prognosis. “For the worst kind of tumour, glioblastoma, patients only survive for about one year even with aggressive treatment after the initial diagnosis,” he said.

So far, there are not many drugs to treat malignant brain tumours. Often, temozolomide combined with radiation is used, but patients can react very differently to it. “Some patients may respond well, and some patients not at all,” he said. A lot of research that has found malignant brain tumours behave differently under the same diagnosis and grading, which makes it difficult for clinicians to determine tumour malignancy precisely. Worse still, the variations among patients have challenged the development of targeted treatment for individuals.

Targeting brain tumours

The first-line treatment includes surgery to safely remove as much tumour as possible, followed by radiation and drug treatment. But visualising the tumour margin accurately for effective surgery is a big challenge for surgeons, and the techniques used are flawed.

“Currently, if we suspect a patient has a tumour, we send them for an MRI scan, analyse the images to see where the tumour is, and use the scan images to guide any surgery,” Dr Li explained. Neurosurgeons use scan images to detect tumour margins. The process involves injecting a contrast agent such as gadolinium into the body to make the tumour mass relatively brighter than the surrounding area, but he said this method is not very accurate. Other methods have been developed to visualise tumours. “For instance, surgeons can use fluorescence agents to show where a tumour is during microscopic surgery. The tumour ‘glows’ so they can differentiate it from normal brain tissue. But still, this technique is not perfect.”

MRI scan images are used to guide radiotherapy and monitor whether a tumour has come back after treatment. “So imaging is the mainstay for patient management,” he said.

Dr Li seized an opportunity to develop imaging techniques at the University of Manchester as a visiting scientist to work on MRI data collected from glioma patients. This also involved learning to program for medical image analysis, along with new research skills, such as understanding the biophysics behind various MRI techniques and MRI data modelling, which were different from the skills he learned during medical training and in his previous research, such as tissue handling.

“From clinical practice, we’re used to just using our eyes to evaluate medical images, but MRI scans can be quantitatively analysed and provide more in-depth information. That inspired me a lot,” Dr Li said. It led him to pursue a PhD in imaging research so he could investigate the best way to analyse numbers from scan images. Traditional statistical models could be used, but he realised he needed to master machine learning to better analyse these images. “Then this beautiful thing started.”

Moving to machine learning

Dr Li began learning traditional ML models and then gradually progressed to the most state-of-the-art deep learning architectures. He believes deep learning models are currently the most powerful tools to analyse MRI scans. For instance, deep learning has shown great potential in helping clinicians identify the tumour mass, predict the diagnosis and prognosis of patients, and delineate the tumour for treatment planning, etc.

He is using his clinical experience, plus knowledge of AI and data science techniques, to translate these models into real-world clinical settings. “This challenge is huge because lots of the models are developed based on standard or benchmark data sets,” he explained. However, clinical settings present more challenges, such as variations in diseases and variations in imaging protocols. “You even get different scanner settings, and so on, so there are lots of challenges for models being made useful in hospitals,” he added.

Dr Li’s other interest is developing more trustworthy AI for small clinical datasets. “We know that lots of deep learning models can work on medical datasets and can help the clinician’s decision making, but we cannot yet completely trust them in making clinical decisions,” he explained. This is a huge barrier to AI tools being adopted in clinical settings, and many factors need to be considered. For example, doctors must understand AI reasoning is sound and why it works.

Dr Li wants to explain how an AI model makes a decision and more importantly, whether the decision corresponds to the clinical domain knowledge. This is vital if an AI model could be safe and effective in a real-world clinical setting, he explained. “For the tasks that AI models are probably good at, we need to evaluate how much confidence medical professionals need to rely on a clinical decision made by AI to be adopted in clinical practice. And we also need to understand what AI can know and what it cannot.”

As well as harnessing the power of AI for his work, Dr Li supervises students from both clinical and STEM backgrounds, including mathematics, computer science and engineering, and helps them step out of their comfort zones, learn cross-disciplinary knowledge and develop the projects of their interests.

Setting his mind to advancing precision medicine using AI

Dr Li has set his mind on advancing precision medicine for neurological diseases using AI. He hopes his work will lead to software or a product that can help clinicians better monitor patients with benign or less malignant brain tumours to predict whether they need surgery or not. “This could enhance monitoring efficiency to provide the most relevant follow-up plan for individual patients. It could also reduce costs for patients or healthcare systems such as the NHS.”

While a lot of research and clinical validation is needed to realise trustworthy AI solutions and it may be years before AI tools are widely used in hospitals, just think of the possibilities.