The method offers a portable, low-cost and safe alternative to X-ray and MRI scans
Uday Khankhoje’s team at IIT Madras is interested in developing a way of detecting breast cancer using microwaves – or radio frequency (RF) waves, as they are called. While several groups have worked on this in Europe and the US, and even made working hardware for this purpose, Dr Khankhoje’s group uses the very popular method of “deep learning” for this. The method not only addresses a mathematical challenge, it also increases the range of the permittivity observed, where permittivity, the square of the refractive index of a material, is the characteristic that distinguishes cancer tissue from normal tissue. Further, this offers a portable, low-cost and safe alternative to X-ray and MRI scans already available for detecting cancer tissue.
In their method, what Dr. Khankhoje’s team would do is to surround the patient with RF transmitters and receivers and collect the waves that bounce off the tissues. Analysing the waves reflected by the tissue, they would reconstruct the type of tissue, or the permittivities of the tissues, that scattered the waves.
This is a classic example of what are known as inverse scattering problems. Other examples of inverse scattering problems are the following: detecting buried landmines using ground penetrating radars; archaeological missions for detecting buried artefacts and so on. These are “inverse” problems because you observe the way waves scatter off an unknown object and reconstruct what it is made of, its shape and other characteristics. The innovation used by this group in solving the inverse scattering problem is “deep learning,” which is a popular technique involving neural networks. Their article has been published in the journal IEEE Transactions on Computational Imaging.
“A neural network is something that learns a relation between input and output just by looking at data,” explains Dr Khankhoje. If you have pairs of numbers (1,1), (2,4), (3,9), (4,16) and so on, a human intelligence would guess that this is a series of numbers and their squares. A machine, on the other hand, “learns” this series, and when given a number as input can produce an output that is the square of that number without having figured out that the relation between them is “square of”. Deep learning, simply, is such a learning process made up of a huge number of “neurons.”
A neural network thus has to be “trained” on data. “We generate our own input or output training data because we know the physics of the problem. This data is used to train the network for inputs it is yet to see,” says Yash Sanghvi the first author of the paper.
He further explains that by this learning, the algorithm positions the analysis approximately in the correct region of the solution. Then existing physics-based algorithms take over, refine the result and arrive at the correct answer. “This exciting new framework of combining physics and machine learning has a very bright future, and in my opinion, it is important to do both,” he adds.
The group is yet to work with the actual biological data. “More work needs to be done, including getting biological samples, building a hardware setup and running trials. That is the direction in which we are heading,” says Dr. Khankhoje.