impact the early detection and treatment of cancer.
The team at Mass General Brigham used the technology that supports foundation models to find new biomarkers for cancer imaging. These biomarkers could change the way we identify patterns in radiological images, leading to better detection and treatment of cancer.
The impact of early detection and treatment of cancer is significant.
A research team used a dataset of 11,467 images of abnormal radiologic scans to develop their foundation model. This model was capable of predicting anatomical site, malignancy, and prognosis across three different use cases in four cohorts. It outperformed existing methods in the field, especially in specialized tasks with limited data. The results were published in Nature Machine Intelligence.
“Image biomarker studies will benefit from this research.”
es are designed to address more specific research questions, which the researchers believe will lead to more precise and effective investigations,” said Suraj Pai, the first author from the Artificial Intelligence in Medicine (AIM) Program at Mass General Brigham.
Although AI methods have shown improved effectiveness, there is still a fundamental question about their reliability and explainability (the idea that an AI’s responses can be understood in a way that “makes sense” to humans). The researchers showed that their methods were consistent across different readers and variations in data collection. The foundation model also revealed consistent patterns.The study found strong links to underlying biological processes, particularly related to the immune system.
“Our results show the effectiveness of foundational models in the field of medicine, especially when dealing with limited data for training deep learning networks. This is especially true when it comes to identifying reliable imaging biomarkers for cancer-related applications,” explained senior author Hugo Aerts, PhD, who is the director of the AIM Program.
