Scientists have taught artificial intelligence to “explain” their decisions

Bioinformaticians from the Ruhr University in Bochum have developed a neural network that can “explain” the course of their decisions. published research Medical Image Analysis.

The study’s authors set out to build the neural network as usual and fed the AI ​​multiple images of microscopic infrared tissue, some containing tumors and others containing only healthy tissue. AI is often inductive, meaning it goes from the specific to the general: using the training data, the AI ​​creates a general model in which it evaluates all other observations provided to it.

At the same time, neural networks are often a black box: it is not clear what features the algorithm chooses from the training data, and it can arrive at wrong conclusions due to inductive logic. The scientists decided to take a different approach: They taught artificial intelligence the deductive method used in scientific knowledge. The neural network they developed uses induction to classify a tissue sample based on the presence of a tumor and also deductively generates a microscopic image map of the tissue. Based on this, a person can then check the accuracy of the model with a special dye, such as molecular methods, for example, histological staining of tissue.

This approach will allow the development of biomarkers with which AI can identify tumor subtypes – these need to be known for effective treatment. In addition, the transparency of the algorithm will increase the trust in it among doctors and patients.

Artificial intelligence can be taught to recognize the signs of a cancerous tumor in a tissue image for a relatively long time. However, exactly how AI decides whether a tumor is in front of it is still often a mystery – unlike human experts, they can’t explain their decision.