A team from France’s Institut Curie is currently working on an artificial intelligence that could be capable of detecting cancer of unknown primary origin. Results are promising.

Tested on 48 tumours, the algorithm detected 79% of them. Treating cancers where the primary tumour is not found is a major challenge, and one that researcher Sarah Watson’s team is looking at in its work that concerns tackling cancers that are discovered only once they have metastasized, i.e., the disease has already spread to other tissues before being identified.

Such cancers are often diagnosed at an advanced stage. In France, there are about 7,000 such patients, or 2% to 3% of cancer cases. The results of this new work on the subject are published in the Journal of Molecular Diagnostics.

In order to find the primary tissue of a case of cancer, and thus propose the best possible treatment, doctors have to perform a battery of tests. The Institut Curie explains that specialised doctors “would run comprehensive medical imaging of the entire body using a scanner and PET scan”, then “conduct detailed analysis via microscope (anatomic-pathology) of samples of these metastases on the lookout for clues as to their origin.

"More recently,” doctors “would draw on molecular biology to spot mutations, particular genetic features that would point to a given organ”, but despite all these approaches lack of identification often led to patients being “treated via non-specific broad-spectrum chemotherapy”.

The deep-learning approach

Researchers at the Institut Curie have now developed a deep-learning artificial intelligence tool to “sequence all genes expressed in a tumour”, outlined Watson. Subsequently, the AI “draws up a diagnostic classifier based on expression profiles of over 20,000 tumors and normal tissues”.

Early results are promising. The scientists submitted 48 tumours of unknown origin, and in 79% of the cases, the tissue of origin was discovered. Of the 11 patients diagnosed, eight have already received treatment.

 - AFP Relaxnews