Researchers at Ben Gurion University in the Negev have developed a computational method that allows them to “reverse engineer” decisions “of artificial intelligence by dividing medical images into components with different clinical interpretations important to artificial intelligence. Understanding the decision-making mechanisms of artificial intelligence models is the key to deciphering biological processes and medical decisions.
The study results were published in the journal Nature Communications.
Deep learning uses artificial neural networks, which are an artificial intelligence-based computing method that can directly learn patterns from data by imitating the learning process of the human brain. The main disadvantage of using this artificial intelligence-based approach is the inability to decipher the reasoning behind neural network decisions.
This limitation stems from the fact that the training process of the network is automated directly based on the data without human intervention. This shortcoming poses a major obstacle to widespread application in fields such as biology and medicine, where interpretation is as important as a machine’s ability to make correct decisions.
Under the guidance of Professor Assaf Zaritsky from the Department of Software and Information Systems Engineering at Ben Gurion University of the Negev, doctoral student Oded Rotem developed a computing method called DISCOVER that reverses artificial intelligence by decomposing images into semantic content. A meaningful part of artificial intelligence through which it makes decisions.
Researchers, working with Israeli startup AIVF, have shown that the technology can characterize the characteristics of in vitro fertilization (IVF) embryos that are most important for artificial intelligence to make decisions about the visual quality of the embryo.
To ensure that the technology can be applied in areas other than IVF, the researchers demonstrated an interpretation of artificial intelligence decisions on MRI imaging of the brains of Alzheimer’s patients, and even explained how artificial intelligence distinguishes between dogs and cats and between men and women in images captured by standard cameras.
The research team used a rich database of thousands of embryos collected by AIVF. Using a light microscope to image the embryos, the company’s embryologists examine and rank each embryo based on several characteristics, such as embryo size and the chain of cells surrounding the embryo in the early stages of development (clinically called trophectoderm).
Researchers have shown that artificial intelligence can successfully predict embryo quality like human experts, but artificial intelligence does not provide researchers with clues as to which embryo characteristics lead to successful predictions.
“Deep learning can identify hidden patterns that the human eye cannot detect in biomedical images. However, this is not enough-in order to make clinical or scientific decisions, we must decipher the mysteries of artificial intelligence recognition and interpret biological or clinical outcomes. Explain the importance of it and decide the next step of treatment or research based on the explanation,”Professor Zaritsky explained.
DISCOVER’s interpretability mechanism relies on “deepfake” generation of artificial intelligence, which, for example, can replace one person’s face in an image with another person’s face. More specifically, the second neural network can create a synthetic image of the embryo in a controlled manner.
The creation of images is based on the definition of certain components in the network, so each component is important on the one hand in predicting embryo quality and on the other hand it encodes meaningful parts of the image. Each such component encodes unique parts of the image, assuming they will be transformed into clear and unique interpretable attributes.
Gradually changing these components, one at a time, can generate images of embryos, each of which differs from the real image in an attribute important to the AI decision-making process.
As a result, the same embryo can be shown to experts in multiple ways, so that in each image, one attribute is artificially “zoomed in” while the rest of the image remains unchanged. This approach allows experts to explain the operating patterns of artificial intelligence and provide objective measurements to demonstrate the importance of each attribute in decision-making.
By creating a series of “fake” images of embryos that had never existed, researchers were able to identify changes in the size of the embryo and the chain of cells around it based on the clinical decisions of embryologists.
In addition, researchers were able to identify a new characteristic that artificial intelligence identified as an important indicator of embryo quality without human guidance, namely the specific structure of the embryo lumen, which contains nutrients from the interior of the cell, clinically described as “blastocyst density.”
“Embryologists are very aware of the importance of certain biological characteristics in determining embryo quality, but the human eye’s ability to accurately measure and evaluate these characteristics is often limited,” explains Daniela Gilboa, CEO and clinical embryologist of AIVF.
“A typical example is blastocyst density, a characteristic that is very important for embryo quality but is not widely used clinically because it is difficult to quantify when visually inspecting embryos in the laboratory. Now, with the help of DISCOVER’s visual interpretation, important biological characteristics can be identified and analyzed more accurately and objectively.
“As a result, we can improve the process of selecting embryos for successful implantation in the uterus, increasing the chances of success in fertility treatment.”
Oded Rotem, a doctoral student who conceived and developed the method, pointed out: “DISCOVER’s ability to identify and artificially amplify image patterns that are important to artificial intelligence can be applied to other areas of biological and medical imaging where artificial intelligence is widely used.”
Dr. Galit Mazuz Perlmutter of BGN, a commercialization company at Ben Gurion University in the Negev, also pointed out the inherent potential of DISCOVER, saying: “The technology developed in Professor Zaritsky’s laboratory has transformative implications for various medical applications.”
There are links in the video below this video. If you are interested, you can open it and have a look.
Thank you for watching this video. If you like it, please subscribe and like it. thank
Original text:https://medicalxpress.com/news/2024-09-method-ai-doctors-decipher-medical.html
More information: Oded Rotem et al, Visual interpretability of image-based classification models by generative latent space disentanglement applied to in vitro fertilization, Nature Communications (2024). DOI: 10.1038/s41467-024-51136-9
More information: Oded Rotem et al., Achieving visual interpretability of image-based classification models through generative latent spatial unwrapping applied to in vitro fertilization, Nature Communications (2024). DOI：10.1038/s41467-024-51136-9
Journal information: Nature Communications
Journal information: Nature Newsletter
Oil tubing: