Artificial intelligence (AI) can identify information about breast cancer and patients' prognosis in digital images of tumour tissue. This is shown in a new doctoral thesis from Karolinska Institutet. The findings suggest that tissue samples already used in healthcare contain more information than is currently used, and that, in the future, this information could complement existing methods for assessing the disease.
When breast cancer is diagnosed, tissue samples from the tumour are examined to provide information that can help doctors assess the disease and choose treatment. Today, these tissue samples can also be scanned and converted into digital images.
In his thesis, Nikolaos Tsiknakis at the Department of Oncology-Pathology investigated whether AI can identify patterns in these images that provide additional information about the tumour and the patient's prognosis.
"Our findings show that these images contain a wealth of information that can be difficult to detect by eye. With AI, we can identify patterns that could eventually give us a clearer picture of the disease," says Nikolaos Tsiknakis.
AI identifies patterns linked to prognosis
The researchers developed AI models that analysed digital images of breast cancer tissue. The models examined both small details within the tumour and the broader structure of the tissue.
The models could estimate certain cellular and molecular characteristics of the tumour from the images. The researchers were also able to use AI to divide patients into groups with different prognoses. The findings were confirmed in independent cohorts.
One potential future benefit is that more information about a tumour could be obtained from tissue samples already collected and analysed in practice. This could complement the methods currently used in healthcare.
Further testing is needed before AI can be used in clinical care
The thesis also highlights important challenges. When the researchers compared different AI models developed for the same task, the models did not always produce the same results. These differences became more apparent when the models were tested on tissue samples from other patient groups.
This underlines the importance of thoroughly testing AI models across different patient groups and healthcare settings before they can be used to support decisions about care and treatment.
Overall, the thesis shows that digital images of tissue samples may contain more useful information than is currently being used. Further research is needed to establish how AI can best be used in everyday clinical practice.
The dissertation takes place on 13 November 2026 in CCK Lecture Hall, R8:00, Visionsgatan 56, Solna
