Researchers at Ben-Gurion University of the Negev have developed a new method, called DISCOVER, to improve understanding of how artificial intelligence (AI) analyzes medical images. DISCOVER allows scientists to “reverse engineer” the AI’s decision-making process by breaking down medical images into key components that are important for the AI’s conclusions. This method aims to build trust in AI’s use in medicine, where understanding the rationale behind decisions is crucial.
Led by Prof. Assaf Zaritsky and doctoral student Oded Rotem, the team showcased how DISCOVER can help doctors interpret AI-driven assessments in areas like in vitro fertilization (IVF). By identifying which parts of an image the AI deems important, the system can provide more transparency, addressing the challenge of AI’s often opaque decision-making process. Their findings, published in Nature Communications, hold promise for various medical applications, including embryo selection and brain scans.
Dr. Assaf Zaritsky’s computational cell dynamics lab team. (Courtesy)
Zaritsky told The Media Line that AI is adept at identifying patterns relevant to clinical decisions. “The problem is, it’s really hard to understand how the machine reaches its decision,” he said. “And in medical decisions, we care about that a lot.”
We really need to understand why, not just receive a diagnosis
He emphasized that understanding the reasoning behind AI decisions is essential. “We really need to understand why, not just receive a diagnosis. It’s not enough to say this embryo is high quality, or this tumor will be more aggressive; we need to catch errors and ensure the machine didn’t use irrelevant information for its decision,” he said. He gave an example where AI mistakenly identified a dog as a wolf due to snow in the image. “This is the kind of error we really want to avoid.”
To address such issues, Zaritsky’s team developed DISCOVER, a method designed to make AI decisions more transparent. “Our goal with DISCOVER is to bridge the gap between AI and the ability to interpret AI,” he explained. DISCOVER takes a two-step approach: “First, we design an AI to explain the AI, creating a disentangled representation. Then, a human comes in to interpret it, and once we have the interpretation, we can say for each embryo or brain why the decision was made.”
This holiday season, give to:
Truth and understanding
The Media Line's intrepid correspondents are in Israel, Gaza, Lebanon, Syria and Pakistan providing first-person reporting.
They all said they cover it.
We see it.
We report with just one agenda: the truth.
Our goal with DISCOVER is to bridge the gap between AI and the ability to interpret AI
He credited Rotem for finding “a really creative way to disentangle these explanations,” noting that Rotem trained AI to adjust specific parts of an image, allowing humans to interpret each change. This collaboration between human expertise and AI aims to reduce uncertainty in decision-making.
A panel of fake (computationally generated) IVF human embryos. (Courtesy)
Rotem explained that DISCOVER was developed with a focus on in vitro fertilization (IVF) and aimed to improve the success rate of embryo selection. He emphasized the importance of building trust in AI decisions. “What happens when there is disagreement between an expert clinician and the AI model? Do we trust the AI model?”
Rotem added, “Our tool helps decrease the uncertainty about what the AI is doing in order to predict results. In this way, we are bringing the AI and the human closer together.”
Zaritsky also discussed the ethical concerns surrounding AI in medicine, particularly the need to avoid bias. “If a machine is trained on images of only white people, it may not know how to respond to people of other colors,” he warned, stressing the importance of ensuring AI decisions are not influenced by irrelevant factors.
DISCOVER has been tested in various clinical fields, including IVF, and has received positive feedback from medical collaborators. “We worked on really hard problems from the medical domain, interpreted them with clinical collaborators, and validated them,” Zaritsky said, adding that an IVF professional recently inquired about using DISCOVER for a case from five years ago.
Zaritsky is optimistic about the technology’s potential. “We’re finalizing a patent and working with our collaborators in the IVF domain to integrate it into their products. In principle, this could be expanded to other domains pretty naturally.”
He noted that while AI can perform as well as, or better than, human experts in many cases, a human is still needed for the final decision. “You need the reasoning behind the machine’s decision to actually make a call with open eyes.”
Rotem concluded that the project exceeded expectations, providing new insights into embryo morphology that were valuable to embryologists.