Regina Barzilay was 42 when a preventive breast cancer screening showed some high-density spots on her mammogram. Radiologists didn’t know whether the spots were normal or cancerous. Two years and three ambiguous clinical tests later, Barzilay was finally diagnosed with breast cancer in 2014. During the course of her ultimately successful treatment, she started wondering whether artificial intelligence (AI) wouldn’t do a better job in diagnosing the disease. Actually, she was quite sure it would.
At that time, Barzilay, an Israeli-American professor at the Massachusetts Institute of Technology (MIT) in the USA, specialised in machine learning for natural language processing, a branch of AI. Although she knew nothing about AI in healthcare, she decided to change course in her career and aim to improve clinical care with the help of AI.
Within four years, she developed an AI program that could diagnose breast cancer and accurately assess the risk of future breast cancer. The program was trained on 32 000 mammograms from women of different ages and ethnicities. It predicted who would be diagnosed with breast cancer within five years of taking the mammogram, and performed better than doctors had previously done.
Computers for medical diagnosis
Trying to use computers for detecting breast cancer on mammograms is not new. It has been tried since the 1990’s. At that time it was called computer-assisted diagnosis (CAD). CAD failed because it didn’t work well enough. It was time-consuming and made too many mistakes. Over the decades things have changed radically, however. Especially since the breakthrough of an AI-method called deep-learning in 2012, computers have become spectacularly better in image recognition. Barzilay and may others, also in Europe, investigate the potential of deep learning for medical diagnosis.
Nowadays, Barzilay is a distinguished professor for AI and health at MIT and the AI faculty lead for the Jameel Clinic, the MIT centre for machine learning in health. She is fully devoted to improving healthcare with AI, especially medical imaging. In her opinion, we are paying too little attention to how weak diagnostics and predictions are at the moment. “This has a terrible effect on the wellbeing and mortality of patients”, Barzilay says in an online interview. Her personal history with the diagnosis of her breast cancer is only one of many examples.
Professor Regina Barzilay, MIT: “It is very important that we bring to public understanding that AI can make a huge difference in healthcare.”
Read the full interview
Most AI-in-healthcare specialists and many radiologists agree that AI will start to automate the rote and tedious tasks in radiology over the next few years. As the American professor of radiology Curtis Langlotz of Stanford University formulates it: “AI won’t replace radiologists, but radiologists who use AI will replace radiologists who don’t.” But in the longer run AI quite probably will also be able to do diagnostics and predictions for disease and treatment which go beyond human’s capacities.
AI for prediction and treatment
While artificial intelligence is gradually being introduced in hospitals to help radiologists make diagnoses, Professor Barzilay is already investigating future AI applications. She is developing AI technology for things that human radiologists cannot do, especially in imaging breast and lung cancer:
Professor Regina Barzilay: “We want to answer questions like: how much earlier can you predict that somebody is going to develop cancer? For breast cancer, we already demonstrated that you can do that quite well, and we validated it with data from 10 hospitals. Or questions like: can we determine exactly who really needs preventive screenings and who does not? Can we predict how well a patient will respond to a certain treatment?”
Hugo Aerts, a Dutch professor with joined appointments at Maastricht University (Netherlands) and Harvard University (USA), uses AI in the field of personalised cancer treatments. His aim is to better predict which cancer patients will benefit from immunotherapy and which ones will not. Immunotherapy does not act directly on the tumour, like chemotherapy, but rather triggers the immune system to recognise and destroy cancer cells. However, the problem is that it is expensive and that it only works for some patients.
Professor Hugo Aerts: “At present we are not good at predicting for whom it will work and for whom it won’t, says Aerts in an online interview. In 2020, he received a five-year European Research Council grant to solve that problem. ‘Our AI technology can quantify the characteristics of tumours on medical imaging better than doctors do. After testing it on independent data, we hope to show that there is a significant improvement in predicting the value of immunotherapy for a patient. If that is the case, then this could be the basis of starting clinical trials, hopefully by 2025.” – Read the full interview
Data for the public benefit
Currently, an estimated 90% of medical data is not shared because of data privacy concerns. If we manage to better balance the benefits and the risks of data sharing, we might make a huge step forward in healthcare. Imagine if we could integrate medical images and data from patients all over the world, of all genders, ages, and ethnicities and at all possible length scales, from genetic and cell data to all possible imaging techniques – how many years of life lost could we save?
Both Barzilay and Aerts believe that to improve medical imaging, it is crucial that researchers gain access to much larger data sets.
Professor Regina Barzilay: “Of course there are concerns about patient data privacy, but there are so many patients whose disease became so much more severe because it was not detected early enough. We have to find a way to balance data privacy with the need for much larger databases.”
Professor Hugo Aerts: “AI is becoming a kind of polar star towards which many hospitals want to navigate. Sometimes there is too much hype around AI, and some disappointment will undoubtedly follow, but the good thing is that the excitement around AI has finally motivated hospitals to create a 21st-century data environment where data is easily searchable.”
Other European projects in AI for medical imaging
• Artificial intelligence in radiology imaging: more than meets the eye
• Automated diagnosis and quantitative analysis of COVID-19 on imaging
• Deep learning for mammography: Improving accuracy and productivity in breast cancer diagnosis