A new AI tool from the Perelman School of Medicine at University of Pennsylvania called iStar (Inferring Super-Resolution Tissue Architecture) offers clear medical image interpretation. It aims to assist time-constrained clinicians by providing highly detailed views of individual cells and a broader look at gene operations.
This innovation can help diagnose and treat cancers, potentially detecting otherwise invisible cancer cells. iStar may aid in assessing safe margins post-cancer surgery and automatically annotate microscopic images, facilitating molecular disease diagnosis.
A new method called iStar was published in Nature Biotechnology and led by Dr. Daiwei Zhang and Dr. Mingyao Li from the Perelman School of Medicine. iStar, based on spatial transcriptomics, employs a machine learning tool, the Hierarchical Vision Transformer, to interpret tissue images.
It breaks down images, capturing fine details and broader tissue patterns, enabling predictions of gene activities at near-single-cell resolution. Notably, iStar can automatically detect significant anti-tumor immune formations, aiding in patient survival prediction and selecting candidates for precision immunotherapy, potentially revolutionizing cancer treatment.
Li explained, “The power of iStar stems from its advanced techniques, which mirror, in reverse, how a pathologist would study a tissue sample. Just as a pathologist identifies broader regions and zooms in on detailed cellular structures, iStar can capture the overarching tissue structures and focus on the minutiae in a tissue image.”
Researchers tested iStar on various cancer tissues, including breast, prostate, kidney, and colorectal cancers mixed with healthy tissues. iStar effectively detected difficult-to-identify tumors and cancer cells, supporting clinicians in diagnosing hard-to-see cancers. Moreover, iStar proved remarkably fast, completing its analysis in nine minutes compared to the best competitor’s 32 hours, making it 213 times faster. This speed enhances its potential for quick and efficient clinical use in diagnosing various cancers.
Li said, “The implication is that iStar can be applied to many samples, which is critical in large-scale biomedical studies. Its speed is also important for its current 3D and biobank sample prediction extensions. In 3D, a tissue block may involve hundreds to thousands of serially cut tissue slices. The speed of iStar makes it possible to quickly reconstruct this huge amount of spatial data.”
Li and her team are extending iStar’s application to biobanks, where extensive sample collections are stored. They aim to enhance researchers’ understanding of tissue microenvironments, providing valuable data for diagnostics and treatments in the future.