Physics-based AI could boost biomedical imaging and autonomous vehicle sensors

A research team led by institutions including the University of California, Los Angeles (UCLA) and the University of Rochester has unveiled a breakthrough in imaging technology that leverages physics-based artificial intelligence. This advancement promises to enhance both biomedical imaging and sensor systems used in autonomous vehicles. The innovation allows for clearer visualization of details hidden within complex media, such as human tissue or dense fog, by significantly improving image quality and processing speed. The new system employs a combination of existing imaging techniques and advanced machine learning algorithms. Specifically, it integrates a method known as DeepTimeGate, which consists of two distinct phases. The first phase involves an algorithm trained to mathematically reconstruct images from scattered light data. The second stage introduces a novel algorithm developed at UCLA that applies physical constraints to ensure the reconstructed images adhere to the laws of physics. This dual-stage approach effectively reduces visual distortions and enhances clarity. In laboratory tests, the system demonstrated remarkable improvements over current technologies. When applied to standard calibration images obscured by complex media, the new system achieved a signal-to-noise ratio more than double that of its predecessors. Additionally, it produced high-quality images almost instantaneously, completing the process in mere thousandths of a second. These capabilities suggest a substantial leap forward in real-time imaging applications. Traditional methods for imaging through complex media often rely on costly equipment capable of detecting near-infrared light. However, the technique developed by researchers at the University of Rochester offers a more affordable alternative. It utilizes silicon-based cameras, similar to those found in consumer electronics, paired with specialized films that filter light to make it visible. While this method provides cost-effective solutions, it previously suffered from issues such as vignetting, which causes darkened image edges and reduced field of view, along with unwanted artifacts appearing as bright or dark spots. By incorporating the DeepTimeGate framework, these challenges have been mitigated. The physics-based checks introduced during the second stage of the algorithm ensure that reconstructed images maintain accuracy while eliminating distortions. This enhancement opens up numerous possibilities, particularly in medical fields where non-invasive imaging is crucial. For instance, surgeons could benefit from more accurate intraoperative guidance, especially during minimally invasive procedures. Similarly, laboratories analyzing biological samples in opaque environments could utilize this technology without the need for sample preparation steps that alter the integrity of specimens. Beyond healthcare, the implications extend to other industries. Autonomous vehicles equipped with such imaging systems could navigate safely through adverse weather conditions, including heavy rain, fog, and sandstorms. Manufacturing sectors might also find value in applying this technology for inspecting products in challenging environments, such as assessing the quality of items submerged in cloudy liquids or wrapped in frosted materials. Waste management facilities could similarly benefit from enhanced visibility during sorting processes. The research was published in the journal Light: Science & Applications under the title “Hybrid deep reconstruction for vignetting-free upconversion imaging through scattering in epsilon-near-zero materials.” The work was conducted by Hao Zhang and colleagues, marking a significant contribution to the field of optical imaging and machine learning integration.