Hidden cancer: Artificial Intelligence helps diagnose tumors of unknown origin

There are tumors that go unnoticed until they metastasize to other organs. A computer tool based on Artificial Intelligence, which is being developed in China, could help answer this enigma: preliminary results surpass those of pathologists in identifying the origin of metastatic cancer cells that circulate throughout the body in cancer patients.

If the preliminary results, published in the journal Nature Medicine, are confirmed, the tool could be useful to oncologists both to improve the diagnosis and treatment of metastatic cancer and to prolong patients’ lives.

Faisal Mahmood, an expert in Artificial Intelligence applications in health at the Harvard University School of Medicine, highlighted in statements to Nature that this is a «very significant discovery that can be used as an assistance tool».

Three out of every 100 tumors are of unknown origin. This percentage increases up to 5% according to several researchers. The prognosis for patients with this type of tumor is usually poor.

In addition to biopsies, one of the ways currently used to diagnose metastatic cancer is New Generation Sequencing (NGS), which analyzes a large number of patient DNA segments in a massive and parallel way.

Another method of analysis is based on analyzing tumor cells from the patient’s body, abdominal or lung fluid, and trying to determine their origin.

At Tianjin Medical University Hospital in China, around 300,000 patients receive cancer treatment each year. And, of these, around 4 thousand are diagnosed through the analysis of tumor cells. Even so, in around 300 people it is impossible to determine the origin of the metastatic cancer, says Tian Fei, a surgeon specializing in colorectal cancer at this medical center.

For this reason, Li Xiangchun, a bioinformatics researcher at Tianjin Medical University, decided to develop a deep learning algorithm with his team. The objective was for the tool to be able to analyze the images and predict the origin of the cancer. This Artificial Intelligence algorithm was trained with around 30,000 images of cells from 21,000 patients with tumors of unknown origin.

Then, they tested their model on 27,000 images. The result was that there was an 83% chance that Artificial Intelligence would accurately predict the origin of the tumor.

Furthermore, there was a 99% chance that the origin of the tumor was included in the first three predictions of the model.

This reduction to three possible origins could be useful, according to Mahmood, as it reduces the number of complementary tests, which tend to be intrusive for patients. Predictions were limited to twelve common sources of cancer, such as lung, ovarian, breast and stomach.