Ductal carcinoma in situ (DCIS) represents a pre-invasive form of breast cancer that can progress to more dangerous stages of the disease. This type of cancer accounts for approximately 25 percent of all breast cancer diagnoses.

Due to the complexity in accurately determining the type and stage of DCIS, patients often undergo unnecessarily intensive treatments. To mitigate this problem, an interdisciplinary research team from MIT and ETH Zurich has developed an advanced AI model. This model enables the identification of different stages of DCIS using simple and accessible breast tissue images. The research has shown that both the condition and arrangement of cells within a sample are crucial for accurately determining the stage of DCIS.

Given the availability of these tissue images, researchers have created one of the largest databases of its kind, which was used to train and test the AI model. When comparing the model's predictions with pathologists' diagnoses, a high level of concordance was found.

In the future, this model can help doctors more efficiently diagnose simpler cases without the need for complicated tests, allowing them more time for detailed analysis of cases where it is difficult to predict whether DCIS will become invasive.

"We have laid the foundation for a better understanding of the importance of spatial cell organization in diagnosing DCIS. We have now developed a technique that can be widely applied. Further research and collaboration with hospitals will be key steps in implementing this model in clinical practice," said Caroline Uhler, professor at the Department of Electrical Engineering and Computer Science (EECS) and the Institute for Data, Systems, and Society (IDSS). She is also the director of the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard and a researcher at MIT's Laboratory for Information and Decision Systems (LIDS).

Combining Images and Artificial Intelligence
Between 30 and 50 percent of DCIS patients develop invasive cancer. However, researchers still do not know which biomarkers to use to predict this transition. Techniques such as multiplex staining or single-cell RNA sequencing can help determine the stage of DCIS, but these methods are too expensive for widespread use.

In previous research, scientists have shown that an inexpensive technique known as chromatin staining can be as informative as more expensive methods. For this study, researchers hypothesized that combining this technique with an advanced machine learning model could provide similar information about cancer stages as more expensive methods.

First, they created a dataset containing 560 images of tissue samples from 122 patients at three different stages of the disease. This dataset was used to train an AI model that learns the representation of each cell's state in the tissue sample image, and based on that, infers the cancer stage of the patient.

However, not every cell shows signs of cancer, so researchers had to find a way to meaningfully aggregate them. They designed a model that creates clusters of cells in similar states, identifying eight states that are important markers of DCIS. Some cell states are more indicative of invasive cancer than others. The model determines the proportion of cells in each state within the tissue sample.

Importance of Organization
"In cancer, cell organization also changes. We found that simply having the proportion of cells in each state is not enough. You also need to understand how the cells are organized," explains Shivashankar.

With this insight, the model was designed to take into account both the proportion and arrangement of cell states, which significantly increased its accuracy. "It was interesting to see how important spatial organization is. Previous studies have shown that cells close to the milk ducts are important. However, it is also important to consider which cells are near other cells," says Zhang.

When they compared the results of their model with samples evaluated by pathologists, the model showed a high level of concordance in many cases. In unclear cases, the model could provide information about tissue sample features, such as cell organization, which pathologists can use in decision-making.

This versatile model can be adapted for use in other types of cancer or even neurodegenerative conditions, which is one of the areas researchers are currently exploring. "We have shown that, with the right AI techniques, this simple stain can be very powerful. There is still a lot of research needed, but we must consider cell organization in more of our studies," concludes Uhler.

This research was partially funded by 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 the Simons Investigator Award.

Source: Massachusetts Institute of Technology