AI cancer diagnosis

A team of researchers at the

has designed a computer program that uses artificial intelligence (AI) to analyse the features of ultrasound images to help doctors predict earlier whether or not a woman's cancer has metastasized.

Knowing whether or not a woman's cancer has metastasized directly affects how her doctors will approach treatment, which may, in turn, influence the outcome of that treatment.

Determining whether a tumour has metastasized is not always straightforward, however. Radiologists often start by using diagnostic ultrasound to non-invasively probe the nearby lymph nodes - tissues where cancer cells first migrate once they metastasize. However, in the early stages of cancer, lymph nodes often appear completely normal, even if the cancer has metastasized.

To judge the new system’s effectiveness, the team retrospectively analysed the diagnostic ultrasounds of 50 women with suspected breast cancer who all had lymph nodes that appeared normal in the ultrasound, suggesting that their cancers had not metastasized.

All 50 women later underwent surgery to remove their cancers and axillary lymph nodes and tissue biopsies of the lymph nodes revealed that 20 of them had metastatic cancer and 30 of them had cancer that remained localised at the time of surgery.

The pilot study was intended to determine if the computer would have accurately identified the 20 metastatic cases based on analysing the ultrasound images of the tumours.

The program performed promisingly well, according to medical physicist Karen Drukker, a research associate and assistant professor in the department of radiology at the University of Chicago.

'We discovered that a computer analysis of breast ultrasound could potentially predict with promising accuracy which patients had metastasis and which did not,' she said.

Next, the researchers plan to start a study in which several radiologists will use the computer program to see if it enhances their ability to diagnose metastasis, again based on retrospective cases for which the answer can later be revealed.