A new AI tool developed at the University of Cambridge could dramatically speed up and refine cancer diagnostics. The system, called SMMILe (Superpatch-based Measurable Multiple Instance Learning), analyzes complex digital pathology images in about one minute, compared to up to 20 minutes needed by an experienced pathologist.
Beyond detecting cancer, SMMILe can map exactly where tumor lesions are located and estimate how aggressive different regions are. This is an essential step toward personalized treatment.
Detailed tumor maps
Most AI tools for pathology require large datasets of meticulously annotated slides, which are costly and time-consuming to produce. SMMILe breaks this barrier. It learns from simple, patient-level labels, such as cancer type or grade, without needing region-by-region annotations. Despite this limited input, the model generates detailed “spatial maps” that show tumor boundaries, subtypes, and their proportions.
“Cancer isn't uniform,” says Dr. Zeyu Gao from the Early Cancer Institute at Cambridge. “Our model doesn’t just confirm cancer; it reveals its internal landscape, highlighting more aggressive subtypes that could guide tailored treatment.”
Strong performance across cancer types
The researchers, who’s study was published in Nature Cancer, validated SMMILe across 3,850 whole-slide images spanning lung, kidney, ovarian, breast, stomach, and prostate cancer. When compared with nine leading AI models, SMMILe matched or exceeded their performance in whole-slide classification, and significantly outperformed them in predicting the spatial distribution of cancer subtypes.
Dr. Mireia Crispin-Ortuzar, co-senior author, highlights the breakthrough: “This method acts like sonar for pathology. It reveals hidden patterns without expensive imaging technologies, using only widely available data.”
Toward personalized oncology
The team plans to extend SMMILe to predict molecular biomarkers, helping clinicians understand tumor behavior at a deeper level. This could support highly personalized treatment decisions based not only on what the tumor looks like, but on its underlying biology.
By accelerating analysis and improving diagnostic precision, SMMILe has the potential to shorten the time to treatment and improve outcomes. As Dr. Gao notes, “AI can help ensure patients receive the right therapy sooner, when it matters most.” Dr. Dani Skirrow of Cancer Research UK adds, “These early findings show how AI could empower clinicians with richer insights, helping deliver personalized cancer care faster than ever before.”
