Researchers at the Shanghai Institution of Pancreatic Disease and Alibaba's DAMO Academy built a model called PANDA, trained to spot the cancer on ordinary non-contrast CT scans, the kind done for unrelated reasons every day. They trained it on 3,208 patients from one center, validated it on 6,239 patients across ten centers, then ran it on 20,530 consecutive patients.
In the real-world test, PANDA detected pancreatic lesions with about 93 percent sensitivity and 99.9 percent specificity, and it beat the average radiologist by 34 percent in sensitivity for pancreatic cancer.
A model is only as good as the population it was trained on, and PANDA's data leaned heavily on patients from China, so performance elsewhere still needs confirming. Catching a shadow on a CT scan is also not the same as diagnosing cancer; every flag still needs a radiologist and a biopsy to confirm it. The team is now testing it as routine opportunistic screening.
Read the original study
Cao et al., Nature Medicine, 2023 · doi.org/10.1038/s41591-023-02640-w
A routine scan, read for a cancer nobody ordered it to find.
Source
- Cao, K., Xia, Y., Yao, J., Han, X., Lambert, L., Zhang, T., Tang, W., Jin, G., Jiang, H., Fang, X., Nogues, I., Li, X., Guo, W., Wang, Y., Fang, W., Qiu, M., Hou, Y., Kovarnik, T., Vocka, M., . . . Lu, J. (2023). Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nature Medicine, 29(12), 3033–3043. https://doi.org/10.1038/s41591-023-02640-w