Researchers from Ben-Gurion University have developed a computational method that explains how artificial intelligence systems make clinical decisions by breaking down medical images into meaningful components
Reverse engineering of artificial intelligence. A computational method developed by researchers from Ben-Gurion University of the Negev enables the decomposition of medical images into components of clinical significance. Understanding the decision mechanism of artificial intelligence (AI) models may affect extensive applications in the understanding of biological processes and in the world of medicine. The research findings were published in the prestigious journal Nature Communications..
"Deep learning" using artificial neural networks is an artificial intelligence-based computational method capable of learning patterns of relationships composed of data directly by imitating the learning process in the human brain. The main disadvantage of using AI-based methods of this type is the inability to understand and explain what is behind the neural network's decision. This limitation is due to the fact that the training process of the network is conducted automatically, directly from the data, without human intervention. This shortcoming constitutes a significant barrier to wider use in fields such as biology and medicine where the explanation is just as important as the machine's ability to make a correct decision.
The doctoral student Oded Rotem under the guidance of Prof. Assaf Zaritsky, from the Department of Software and Information Systems Engineering at Ben-Gurion University of the Negev, developed a computational method that allows the AI to be reverse-engineered by decomposing an image into semantically meaningful components through which the AI makes the decision. In collaboration with the Israeli start-up company AIVF, the researchers demonstrated the technology's ability to characterize the features of the fetus that were most significant to the AI in order to make a decision.
To make sure that the technology can be used in contexts beyond the world of in vitro fertilization, the researchers demonstrated an interpretation of the AI's decision also for MRI images in the brains of Alzheimer's patients and even in images taken by a normal camera to interpret how the AI distinguishes between dogs and cats and between men and women .
The research team used a rich image database of thousands of embryos from the IVF process collected at AIVF. The embryos were imaged using a light microscope, and embryo development experts (embryologists) in the company examined and graded each embryo based on a number of features such as: the size of the embryo and the chain of cells that surrounds it in the early stages of development. The researchers demonstrated that the AI succeeds in the task of predicting embryo quality with similar performance to the human expert, but the AI did not offer the researchers any clues as to what embryo features led to the success of the prediction.
"Although deep learning makes it possible to identify hidden patterns in biomedical imaging data that the human eye does not recognize, but to be able to characterize and make clinical or scientific decisions, we must solve the mystery and find out what the AI has detected, interpret the biological or clinical meaning of the explanation, and decide According to the interpretation the next steps in treatment or research," explained Prof. Zaritsky.
In the next step, the researchers developed a new "deep-fake" style method, which allows, for example, to replace one person's face with another person's in a photo. The method, which bears the name DISCOVER, is based on another neural network capable of creating synthetic images of embryos in a controlled manner. The creation of the images is based on the definition of certain components in the network, so that each component on the one hand will be significant in predicting the quality of the embryo and on the other hand will encode meaningful image parts. Each such component encodes unique image parts under the assumption that they will translate into an unambiguous and different feature from component to component. Gradual change of these components, each component separately, allows Creating images of embryos that each differ from the true image in one feature that is important to the AI's decision process. In this way, the same embryo can be presented to the expert in several different ways, so that in each image one feature is artificially "enhanced", while the rest of the image remains unchanged. In this way, the method allows the expert to interpret the operation of the AI and even indicate how important each feature was in the decision.
By creating a series of "fake" images of embryos that never existed in reality, the researchers were able to detect a change in the size of the embryo and in the chain of cells surrounding the embryo - according to the decision made by the embryologist at the clinic. Beyond that, the researchers were able to identify a new feature that the AI recognized as a feature that represents a quality embryo without human guidance - a certain structure of an internal space in the embryo that contains nutrients for the inner mass of the cells, clinically described as "blastocyst density".
"Embryologists are well aware of the importance of certain biological features in determining the quality of the embryo, but the human eye is often limited in its ability to measure and evaluate them accurately," she explained Daniela Gilboa, CEO of AIVF and a clinical embryologist by training. "An excellent example of this is the density of the blastocyst, a feature of great importance in the quality of the embryo that is not widely used clinically because it is very difficult to measure and quantify it with the human eye when examining the embryo in the laboratory. Now, with DISCOVER's visual explanation, such important biological features can be identified and analyzed more accurately and objectively. As a result, we can significantly improve the process of selecting the embryo with the highest chances of successful implantation in the uterus, thereby increasing the chances of success of the fertility treatments."
"DISCOVER's capabilities to identify and artificially amplify image patterns that are important to AI to enable interpretation can also be used in the fields of biological imaging, medical imaging and other fields where artificial intelligence is a filter tool," he noted Oded Rotem, the doctoral student who conceived and developed the method. Dr. Galit Mazuz Perlmutter, from BGN, the commercialization company of Ben Gurion University of the Negev, also noted the inherent potential of the computational method: "The development of Prof. Zaritzky and the laboratory team is of applied importance for various fields in the world of medicine."
The research team included Tamar Schwartz, Ron Maor, Yishai Tauber, Maya Zarfati-Shapiro, Daniela Gilboa and Prof. Daniel Zeidman from the AIVF company, as well as Prof. Marcus Masger from the IVI Valencia Fertility Clinic in Spain.
This research was supported by the Rosetree Trust and by the Israel Council for Higher Education through the Data Science Research Center, Ben-Gurion University of the Negev.
