The structure of SMMILe.

In the field of precise cancer diagnosis and treatment, gigapixel-level digital pathology whole-slide images (WSIs) are considered the highest standard. While existing mainstream AI methods, such as multiple instance learning (MIL), can determine whether cancer is present on a slide at low cost, they cannot tell us where exactly the lesion is located, how the malignancy is distributed, and what the proportion of different subtypes is.

Obtaining this key information often requires pathologists to invest immense effort in manual annotation pixel by pixel, a task that is practically impossible in a clinical setting.

To address this challenge, a team led by Professor Li Chen from the School of Computer Science and Technology and the National Medical Innovation Industry-Education Integration Platform at Xi'an Jiaotong University (XJTU), in collaboration with the University of Cambridge, proposed a novel artificial intelligence framework named SMMILe. The related paper was recently published in the international authoritative oncology journal Nature Cancer.

SMMILe is the first AI system capable of achieving accurate spatial quantification of lesions across an entire slide using only simplified "patient-level diagnostic labels." SMMILe breaks the limitations of traditional weakly supervised algorithms that prioritize classification over localization.

Without the need for expensive manual annotation, it can automatically infer the tumor's specific location, boundary range, and the spatial distribution of different subtypes within the tissue, much like drawing a map.

SMMILe integrates cutting-edge mathematical models, including feature compression, parameter-adaptive processing, and Markov Random Field constraints, enabling it to acutely capture even faint pathological signals.

This approach not only solves the interpretability problem of AI models but also marks a major leap forward in pathological analysis efficiency – a complex tissue slice that might take a human 20 minutes to analyze can be processed by SMMILe to generate a detailed quantitative report in approximately one minute.

By generating reliable spatial quantitative maps, the method provides pathologists with more intuitive tissue structure information, aiding in the quicker and more accurate identification of key areas in complex cases, ensuring patients receive the best treatment plan sooner.

It also provides a powerful tool for researchers to study tumor heterogeneity and explore the relationship between different tissue subtypes and prognosis, immune response, and drug sensitivity.

In the future, this framework is expected to be further expanded to infer the tumor's molecular characteristics, tightly integrating tissue morphology with multi-omics data, thus advancing comprehensive cancer medicine and precision medicine.