Germany: artificial intelligence (AI) can be trained to detect whether or not a tissue image is present tumor. However, until recently, how it makes its decision has been a mystery. A team from Ruhr-Universitat Bochum’s Research Center for Protein Diagnostics (PRODI) is working on a new approach that will make AI decision making transparent and therefore trustworthy.
Researchers led by Prof Axel Mossig describe the approach in the Journal of Medical Image Analysis.
For the study, bioinformatics scientist Axel Mosig collaborated with Professor Andrea Tennafel, Institute of Pathology, oncologist Professor Anke Renascher-Schick of the St. Joseph Hospital of the Ruhr-Universität and Professor Klaus Gerwert, biophysicist and founding director of PRODI. The group developed a neural network, namely AI, that can classify whether a tissue sample contains a tumor or not. To this end, they fed the AI a large number of microscopic tissue images, some of which contained tumors, while others were tumor-free. Researchers explore how artificial intelligence can be trained to detect tumors
“Neural networks are initially black boxes: it is unclear which features the network learns from the training data,” explains Axel Mosig. Unlike human experts, they lack the ability to explain their decisions. “However, especially for medical applications, it is important that AI is capable of explanation and therefore reliable,” adds bioinformatics scientist David Schumacher, who collaborated on the study.
AI is based on false assumptions
So the Bochum team’s explainable AI is based on meaningful statements known to science: on false assumptions. If a hypothesis is false, this fact must be demonstrable by experiment. Artificial intelligence generally follows the principle of inductive reasoning: using concrete observations, i.e. training data, the AI creates a general model based on which it evaluates all further observations.
The underlying problem was described by the philosopher David Hume over 250 years ago and can be easily explained: no matter how many white swans we observe, we can never conclude from this data that all swans are white and that no black swans exist. So science uses what is called deductive reasoning. In this approach, a general hypothesis is the starting point. For example, the hypothesis that all swans are white is disproved when a black swan appears.
The activation map shows where the tumor is found
“At first glance, inductive AI and the deductive scientific method seem almost incompatible,” says Stephanie Schorner, a physicist who also contributed to the study. But researchers found a way. Their novel neural network not only provides a classification of whether a tissue sample contains a tumor or is tumor-free, it also creates an activation map of microscopic tissue images.
The activation map is based on a false assumption, i.e. the activation obtained from the neural network exactly corresponds to the tumor regions in the sample. Site-specific molecular methods can be used to test this hypothesis.
“Thanks to the interdisciplinary structures at PRODI, we have the best prerequisites to incorporate a hypothesis-driven approach into the development of reliable biomarker AI in the future, for example, to be able to differentiate between specific therapy-relevant tumor subtypes,” Axel concludes. Mossig.
