A cloud brain that can recognize breast cancer

If you find a lump in your breast, you could have breast cancer. It could also be nothing - just a benign tumour. You want the answer quickly and when the answer arrives you want it to be right! As her family had been affected by breast cancer 17-year old Brittany Wenger wanted to make a difference. In that situation most people do charity events like a sponsored fun run to raise money for cancer charities. Brittany though was interested in science and particularly computer science, so she decided she could make a more direct difference herself. She created a new way to analyse the results of breast cancer tests and then ran a scientific experiment to check how effective it was. As a result she won Google's 2012 Science Fair and developed a program that could save lives.

One of the least invasive ways doctors can check a lump for cancer is to insert a fine needle into it to get a sample of the cells. They can then run tests on them and based on the results decide if the tumour is benign or not. Deciding isn't quite as easy as that though. There are a range of tests, like how uniform is the shape of the cells, how thick is the clump of cells, do certain cells tend to stick together, and so on. None are completely conclusive alone. Ultimately doctors have to make a judgement call given all the evidence they have. Unfortunately, they don't always get it right. They sometimes decide a tumour is malignant when it is safe. That leads to women having cancer treatment that isn't needed. Worse though is if they decide a lump is benign but it turns out to be cancerous. The chance to save that woman's life might be lost.

If humans aren't good enough at telling the difference, perhaps computer brains could do better. In particular, 'neural networks' - a form of artificial intelligence - may be able to make the decisions more accurately. A neural network is just a computer program that works the way the human brain does. It is made up of a software version of neurons - brain cells - that are connected together in a similar way to the way our own neurons are connected. This kind of computer program is very good at spotting patterns, and that is actually what is needed for diagnosing breast cancer. Given the pattern of numbers that are the results from a woman's breast cancer tests, does it best match the pattern of a cancer or does it instead match the pattern of a benign tumour?

In the 1990s, the University of Wisconsin collected lots of results from cancer tests together with a final definitive answer of whether that person turned out to have a tumour. This provided a large body of data that could be used to base decisions on. It was used to test whether a neural network could accurately predict whether a tumour was cancerous or not. Unfortunately, the results were patchy. The neural networks of the time were just not good enough.

Brittany realised though that now, 20 odd years later neural network technology had improved. Maybe they would be good enough now, especially if tuned specifically for the purpose. So she decided to find out. She took the Wisconsin data and trained three different modern neural network systems with it. That is she gave them some of the test data together with the final decision for those cases. They used it to look for patterns corresponding to tumours being malignant or benign. With each new person's set of test results, they adjusted the weighting they placed on each different test so as to adjust their decision to match the known result for each test. Once they were trained she had them come up with predictions for all the remaining data. She used the results as an experimental control - a baseline to compare against to find out if she could do better.

She then coded her own neural network in Java designing it so that it would specifically put emphasis on avoiding false negatives - she most of all wanted it to avoid saying a cancerous tumour was ok. As with the other neural networks she trained hers on some of the data and then made it predict the rest. Not only was it better at predicting than neural network used in the original 1990s trial, her neural network did substantially better than the modern off the shelf networks too. Overall, hers got it right over 97% of the time. It was even better, 99% of the time getting it right when given a malignant case. That is better than doctors generally manage.

She also showed that the more data a neural network has to work on the more accurately it can potentially become as the patterns emerge more strongly. With a full data set it would be 100% accurate. She therefore went on to create a cloud computing version of her program. That means it is available to anyone to use and at the same time it can collect more results and so improve its accuracy even more.

Brittany has shown that neural networks are a viable way of diagnosing breast cancer based on the quick and simple needle test. If further tests confirm its accuracy, her artificial brain may soon be being used for real, actually saving women's lives. Which just shows, with determination and some computer science skills you really can make a difference.