University trains AI to analyse cancer images

Researchers have developed an AI-based computing model which can count cells from histopathological cancer tumour images.

A team from the University of Jyväskylä state that they have taken the first step towards developing a digital service centre based on artificial intelligence where doctors and pathologists analyse tumour tissue samples visually with the help of software.

The computing model is able to determine the T-cell count in cancer tissue based on nothing but a digital image and with an error margin of a few percent, according to tests. The researchers tested the model against five 523 images of intestinal cancer tumours where previously, the T-cell count of each image was determined by histopathologists. The team found that the model was successful in 90% of cases.

Errors occurred when the model was unable to recognise cells correctly due to potential inaccuracies in the data used to train the neural network. Additional samples will be needed to increase the accuracy of the neural networks and the researchers suggest that results may eventually become more accurate compared to current methods.

The use of AI can assist with labour intensive processes such as counting T-cells and has the potential to speed up similar jobs in healthcare, allowing doctors to spend more of their time researching diseases and treating patients.

Leader of the research, adjunct professor Sami Äyrämö, said: “The training of the neural networks took place in a closed AI computing environment built for analysing social and health care data at the Faculty of Information Technology of University of Jyväskylä. There the researchers have access to a powerful IBM Power 9 AI computer. With the help of neural networks, one sample image can be analysed in seconds in the computing environment, meaning that in one day even thousands of samples can be analysed. However, the results from our pilot do still require validation using an independent dataset.”

Professor Pekka Neittaanmäki, said: “AI based digital pathology offers many new possibilities. Biobanks and hospitals can send their histopathological samples in digital form for analysis to a specialised centre that can provide results very quickly. In the beginning, the centre could serve customers on a national level, and in the future, on an international level.”