A team of researchers has created an artificial intelligence (AI) model that can identify the different stages of ductal carcinoma in situ (DCIS), a preinvasive tumor that may progress to a lethal form of breast cancer.
DCIS accounts for 25% of all breast cancer diagnoses, and up to half of all patients with DCIS may develop a highly invasive stage of cancer. However, no biomarkers have been identified for accurate prediction of which tumors will progress, and it may be difficult for clinicians to determine the stage of the tumor when DCIS is diagnosed. Thus, patients with DCIS may be subjected to overtreatment.
The interdisciplinary team, comprising researchers from the Massachusetts Institute of Technology (MIT) in Cambridge, MA, ETH Zurich in Switzerland, and the University of Palermo in Italy, described their model in Nature Communications.
The team first created a dataset containing 560 tissue samples from 122 patients at three different stages of DCIS. They then used this dataset to train their AI model to identify representative states of cells in an image, which the algorithm then used to infer the stage of the cancer.
The dataset used breast tissue images produced with chromatin staining, which is inexpensive and easy to obtain. One of the objectives was to produce accurate analysis with the AI model using inexpensive imaging that would provide as much information as more costly techniques.
As not all cells are indicative of cancer, they designed the model to identify eight states that are important markers of DCIS. Furthermore, they wrote, "the organization of the identified eight cell states is significantly altered in the different disease stages and phenotypic categories, both in terms of their relative location with respect to the breast ducts and their co-localization with cells from each cell state." Thus, they designed the model with the proportion and arrangement of cell states in mind; this increased the accuracy of the model.
The results using the model were mostly comparable to those produced through evaluation by a pathologist. Where results were more ambiguous, the authors suggested that the model could assist a pathologist in decision-making by providing information about features in the tissue sample.
The authors said in an MIT news release that they are exploring the adaptation of the model for other types of cancer or even neurodegenerative conditions.
