Recently, the Tencent Life Science Laboratory, together with the First Affiliated Hospital of Guangzhou Medical University and the Guangzhou Institute of Respiratory Health, jointly developed the DeepGEM pathological large model, which has made significant progress in predicting lung cancer gene mutations. This model can complete mutation prediction within one minute using only routine histopathological slice images, with an accuracy rate ranging from 78% to 99%, bringing new hope for precision medicine.

QQ20251013-113211.png

This innovative technology stands out by no longer relying on expensive and time-consuming genetic sequencing, but instead using AI to analyze pathological images to identify potential gene mutations. This means that test results that once took thousands of yuan and weeks to obtain could, in the future, be completed in just a few minutes, with costs potentially reduced by several times. This breakthrough will provide lung cancer patients with more precious treatment time, while also offering doctors an efficient and convenient AI-assisted diagnostic tool.

The core technology of DeepGEM lies in its ability to identify "morphological signals" related to gene mutations from ordinary pathological images through AI. Research found that the arrangement of tumor cells, morphological features, and the response of surrounding tissues have a statistical association with certain gene mutations. The DeepGEM model uses a large amount of pathological data to capture these details that may indicate mutations from seemingly ordinary images, significantly improving the accuracy of predictions.

The model adopts a multi-instance learning (MIL) architecture, capable of directly processing entire pathological images without manual annotation of tumor areas. This automated analysis method not only improves diagnostic efficiency but also outputs spatial distribution maps of gene mutations, clearly showing mutation differences in different regions of the tumor, helping doctors quickly identify high-frequency areas.

Considering the morphological differences in pathological samples among different patients, DeepGEM was designed with adaptability in mind, enabling it to process various routine histopathological sections, reducing the application threshold. Currently, the model's prediction accuracy has reached 78% to 99% in multiple data tests, comparable to traditional genetic testing methods, providing doctors with a reference for rapid decision-making.