A newly-published report suggests that artificial intelligence (AI) might help pathologists predict recurrence of mesothelioma.
The report focuses on technology developed by the RIKEN Center for Advanced Intelligence Project (AIP) in Japan. The Center claims the computer program taught itself to recognize key features in images of cancer.
Some of the features were ones that even pathologists did not know about. The findings suggest that AI and pathologists could predict recurrence of mesothelioma better than pathologists alone.
How AI Might Predict Recurrence of Mesothelioma
A computer program usually has to be “taught” to perform a certain task. But this limits how far the learning can go.
Instead of “teaching” it about cancer, researchers let the RIKEN technology learn on its own. They gave the program images from more than 13,000 whole mount pathology slides of prostate cancer. The slides were subdivided into 86 billion smaller images or “image patches”.
As predicted, the AI found features in the images that could help predict recurrence of mesothelioma and other cancers.
Pathologists already knew about some of the features that the AI picked out. These features are already part of the Gleason score for diagnosing prostate cancer.
But some of the features identified by the technology had never been noticed by human scientists before. These were features that involved the stroma, connective tissues that support organs.
“I was very happy to discover that the AI was able to identify cancer on its own from unannotated pathology images,” said study author Dr. Yoichiro Yamamoto. “I was extremely surprised to see that AI found features that can be used to predict recurrence that pathologists had not identified.”
Testing the Accuracy of the Program
Malignant mesothelioma is highly resistant to most cancer treatments. The likelihood of a tumor growing back can impact how doctors treat it. But it is difficult to predict recurrence of mesothelioma.
The new study shows that the RIKEN AI might help.
Once the program had taught itself to recognize signs of cancer, researchers used another 2,275 pathology images to confirm its accuracy. The images came from patients at St. Marianna University Hospital and Aichi Medical University Hospital in Japan.
The group found that the features discovered by the AI tended to be more accurate than the Gleason score. But the best chance to predict recurrence of mesothelioma would probably involve both AI and pathologists.
“Combining the AI’s predictions with those of a pathologist increased the accuracy even further, showing that AI can be used hand-in-hand with doctors to improve medical care,” says Dr. Yamamoto.
The study suggests that the technology might help find other disease characteristics that doctors miss. The more doctors learn about mesothelioma and other hard-to-treat cancers, the better treatment outcomes will be.
Yamamoto, Y, et al, “Automated acquisition of explainable knowledge from unannotated histopathology images”, December 18, 2019, Nature Communications, https://www.nature.com/articles/s41467-019-13647-8
Wilkinson, Jens, “Artificial intelligence identifies previously unknown features associated with cancer recurrence”, December 18, RIKEN website, https://www.riken.jp/en/news_pubs/research_news/pr/2019/20191218_2/index.html