This advancement helps to streamline clinical diagnoses and could extend to other cancers and neurodegenerative diseases.
Ductal carcinoma in situ (DCIS) is a preinvasive breast cancer that sometimes progresses to a more lethal form. Representing about 25% of all breast cancer diagnoses, DCIS often leads to overtreatment due to the difficulty in determining its type and stage. Addressing this issue, an interdisciplinary team from MIT and ETH Zurich has developed a groundbreaking AI model capable of identifying the stages of DCIS from easily obtainable breast tissue images.
This model underscores the importance of both the state and arrangement of cells in diagnosing DCIS. Leveraging the simplicity of acquiring such tissue images, the researchers constructed one of the largest datasets of its kind to train and test their model. When compared to pathologist conclusions, the AI predictions showed significant agreement, promising a new tool for clinicians to streamline diagnoses. Combining imaging with AI, the team addressed the challenge of determining which DCIS tumors might become invasive. Traditional techniques like multiplexed staining or single-cell RNA sequencing, though informative, are prohibitively expensive. The researchers previously found that chromatin staining, a cheaper technique, could be equally informative.
By integrating this stain with a machine-learning model, they aimed to replicate the insights from more costly methods. Their dataset included 560 tissue sample images from 122 patients across three disease stages. The AI model learned to represent the state of each cell, inferring the cancer stage by clustering cells into eight key states. This dual consideration of cell proportion and arrangement enhanced the model’s precision, aligning closely with pathologist evaluations.The team claims that the research shows the power of simple stains when paired with advanced AI techniques, emphasizing the need for further studies on cell organization.
“We’ve taken a crucial first step in recognizing the need to examine the spatial organization of cells when diagnosing DCIS. Our scalable technique now needs a prospective study, collaborating with hospitals to bring it to clinical practice,” says Caroline Uhler, a professor at MIT’s Department of Electrical Engineering and Computer Science and the director of the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard.
