Signs of viruses that can lead to cancer, like the human papillomavirus (HPV) or Epstein-Barr virus (EBV), can be found in tissue samples with costly tests that aren鈥檛 always accurate. But a new study shows that artificial intelligence can hunt down these signs reliably 鈥� and can help teach human doctors to spot them too.
Some cancers can have viral or non-viral pathways, such as gastric cancer or head and neck cancer. In these cases, it鈥檚 crucial to determine whether a virus is involved in order to decide on the best treatment.
Jakob Nikolas Kather at University Hospital RWTH Aachen in Germany and his colleagues trained a neural network using images of tumour tissue samples from The Cancer Genome Atlas. The training images came from 412 people with head and neck cancer, with 12 per cent testing positive for HPV, and 317 people with gastric cancer, with 8 per cent testing positive for EBV.
Advertisement
The neural network accurately identified the presence of a virus in samples with HPV 89 per cent of the time, and 80 per cent of the time for samples with EBV.
The team then reverse-engineered fake tissue sample images from the neural network, using a computer vision algorithm called Deep Dream that let them see what key characteristics in the patterns the AI was 鈥渟eeing鈥� when it found signs of the viruses.
They showed the results to a panel of expert pathologists who described the features. They saw in the HPV-negative images 鈥渁 sheet of small nodules composed of bright, predominantly warm colours,鈥� and in the HPV-positive images 鈥渓arge nests with rounded borders composed of dark, predominantly cool colours punctuated by red dots.鈥�
The EBV-negative images had 鈥渋ll-defined dark green whorls punctuated by blue dots and wisps of yellow鈥� and the EBV-positive images showed 鈥渙verlapping sheets with reticulated patterns in pastel colours.鈥� The team behind the study suggest these patterns could help human doctors better identify the signs of cancer-related viruses.
Reference: bioRxiv, DOI:
