Today, when we visit the website of many companies dealing in consumer goods (e.g. - banks), we are greeted by a chat "representative" on the home page.

We tell this "representative" the reason for our visit (for e.g. - I am looking to get a new credit card) and we are redirected to the page where we would find relevant information. Some companies go a step further and ask a series of questions (for e.g. spending habits, preferences) and try to determine which product would work best for us.

While sometimes a human being is actually at the other side, often this "representative" is actually just a robot that understands the information you provide, interprets it and helps you solve the problem. These bots or systems are designed using a field of computer science called "natural language processing" (NLP) that focuses on computer-human interaction.

Modern NLP systems use machine learning, an approach in which the system learns from existing enterprise-wide data about the potential customer questions and how to respond to them. The system is designed to learn from its experiences and as a result, as the system gets more questions, it gets better and better.

Now, scientists have applied this technology to read mammograms and diagnose potential breast cancer patients. The work was the result of collaboration between Jenny C Chang M.D., Director of Houston Methodist Cancer Center and Dr. Stephen T Wong, Chair of the Department of Systems Medicine and Bioengineering at Houston Methodist Research Institute.

While several advances have been made in our fight against breast cancer, it is still responsible for the death of several women in US every year. To be better prepared, numerous women get mammograms every year and screen against breast cancer.

However, due to a large number of false positives and ambivalent results from mammograms, many questions have been raised regarding the issues of over biopsies, diagnosis, treatment and the huge costs of these mistakes. The researchers have aimed to solve this problem by designing a software, which would help doctors correctly interpret mammograms and reduce costs as well as mitigate the menace of false positives.

According to Tejal Patel, M.D., breast medical oncologist at the Houston Methodist Cancer Center, who is the first author of this paper, a big challenge in mammography research is that a lot of information is buried in notes that are not organized in a pre-defined form. She laments that "despite breast imaging reporting and data system (Bi-RADS) there is variation in reporting terminology."

Dr. Wong agrees and adds that Bi-RADS also does not discriminate enough between patients who have different risks of breast cancer. Besides, a model in which doctors evaluate each mammogram individually is very difficult to scale up efficiently.

So clearly this is a situation where NLP programs, which interpret human interactions correctly would be very helpful and this was the motivation behind the current research.

The group developed a natural language processing software - Methodist Hospital Text Teaser (MOTTI), which learned how to diagnose breast cancer, leveraging all the existing mammographic reports.

As a first step, they tested the software against mammograms showing Bi-RADS category 5 breast cancer i.e. mammograms showing unambiguous indications of breast cancer. The software achieved 99% accuracy of expert clinicians but was many orders faster. In the next phase, the software is going to be tested against more ambiguous cases of breast cancer (Bi-RADS category 4) and the goal is to accurately sub-classify breast cancer risks.

Dr. Wong hopes that if the software is successful against this category, "this will be a valuable disease management tool to aid the clinicians to make better assessment of breast cancer risks, evaluate which mammograms have high risk of malignancy, and identify patients that do not need biopsy."

Similarly, Dr. Patel says that she hopes "to use the software and answer many unanswered questions and also create a predictive tool that can help clinicians make decisions with an abnormal mammogram."

We are still not at the stage where a computer software can largely replace radiologists but this study shows that we are well on our way to a world in which radiologists and other doctors can be heavily supported by artificial intelligence.

Image Credit: http://negativespace.co/photos/coding-stock-photo-4/