Researchers at MIT and ETH Zurich have developed a smart computer that can help doctors determine which early-stage breast tumors are more likely to become deadly. This could spare many women from needless treatment.
Key points
The system concentrates on ductal carcinoma in situ (DCIS), a condition that accounts for about 25% of breast cancer diagnoses. DCIS is a type of tumor that sometimes goes invasive, but often does not. The trouble is, doctors cannot always tell which is which.
To teach a computer to solve the problem, the team fed it with images of breast tissue samples — cheap and easy to make compared to some other tests for breast cancer. It learns what normal cells and their nuclei look like and builds rules around how they are arranged in space.
However, instead of counting different types of cells, the researchers decided to focus on how they were organized — and found this feature much more helpful. According to study author Jochen Steppan, both cell composition and architecture turned out to be important when predicting whether any given DCIS would turn into an invasion.
The algorithm was pitted against what expert doctors thought about the same tissue samples; it agreed with them in many cases, and when things were less clear-cut, it could give them extra information.
This tool is not meant as a replacement for physicians but as an assistive device. It could allow them to work faster in simple cases where the prognosis is obvious so as not to delay complex ones where everything is much less certain.
The team concluded by saying their goal was “not achieving human performance”, but rather providing additional data points for clinicians’ consideration. To realize these ambitions further work needs doing – especially assessing robustness across larger datasets collected from multiple institutions outside its creators’ walls – before being deployed within hospitals as an aid during diagnosis procedures for breast cancers stages I-III+ while ensuring those most at risk receive appropriate care without overburdening limited resources available elsewhere.
