Ductal carcinoma in situ (DCIS), a type of tumor that can develop into an aggressive form of breast cancer, accounts for approximately 25% of all breast cancer diagnoses. DCIS can be challenging for clinicians to accurately categorize, leading to frequent overtreatment for patients. A team of researchers from the Massachusetts Institute of Technology (MIT) and the Swiss Federal Institute of Technology Zurich (ETH Zurich) have set out to improve this situation with artificial intelligence (AI).
The researchers developed an AI model that uses breast tissue images to identify different stages of DCIS. The team’s model found the state and arrangement of cells in a tissue sample are significant markers for determining stages of the tumor. To test the model, the researchers compared its predictions to those from a pathologist, showing substantial agreement in multiple instances. This comparison indicates that the AI system could help streamline diagnostic procedures for less complicated cases in the future, allowing clinicians more time for more complex cases where it is unclear if DCIS will become invasive.
The researchers compiled one of the largest datasets of its kind due to the ease and cost-effectiveness in obtaining tissue images, providing sufficient data to train the AI model. The model was trained on a dataset of 560 tissue sample images from 122 patients at different stages of disease, enabling detailed representation of each cell state in a tissue sample.
The team’s work is not finished, however. They recognized that not all cells are indicative of cancer, so they programmed the model to cluster cells in similar states, and identified eight states as crucial indicators of DCIS. Understanding that the organization of cells is also a marker in cancer, they designed the model to consider both the proportion of cells and their arrangement. This expanded criteria drastically improved the model’s accuracy.
“The interesting thing for us was seeing how much spatial organization matters,” pointed out Xinyi Zhang, the lead author of the paper. He further explained that it’s critical to understand which cells juxtapose each other in addition to those close to the breast duct.
Furthermore, this AI model also presents valuable capabilities for future medical diagnostics, as it could be adapted to identify other forms of cancer or neurodegenerative conditions. The team is already exploring these possibilities.
The research received funding support from multiple institutions, including the Eric and Wendy Schmidt Center at the Broad Institute, ETH Zurich, the Paul Scherrer Institute, the Swiss National Science Foundation, the U.S. National Institutes of Health, the U.S. Office of Naval Research, the MIT Jameel Clinic for Machine Learning and Health, the MIT-IBM Watson AI Lab, and a Simons Investigator Award. A paper detailing their research was published in the Nature Communications journal on July 20, 2021. Finally, the researchers plan to conduct a prospective study in collaboration with a hospital to bring the AI model closer to practical clinical use.