Deep Learning Reconstruction for Low-Dose Dual-Energy CT Pulmonary Angiography: Optimization of Energy Levels and Assessment of Subsegmental Arteries

Pulmonary embolism, a blockage in the blood vessels of the lungs, is a medical emergency that demands swift and precise diagnosis. The current gold standard for finding these blockages is a specialized type of X-ray scan called a CT pulmonary angiography. This procedure involves injecting a liquid dye, known as contrast media, into a vein so that the blood vessels light up brightly on the scanner, allowing doctors to see if a clot is present. However, this life-saving tool comes with a cost. The radiation required to take the pictures can accumulate over a patient's lifetime, and the large volume of dye needed can strain the kidneys, especially in older adults or those with existing health issues. For years, doctors have sought a way to get the same clear pictures while using less radiation and less dye, a goal often called a "double-low" approach.

The challenge lies in the physics of the scan. To make the dye stand out more clearly against the background tissue, scientists can tune the scanner to use lower energy levels. Think of it like turning down the brightness on a flashlight to make a specific color pop; in this case, lower energy makes the iodine-based dye appear much brighter. But there is a catch: lowering the energy also makes the image grainier and noisier, like static on an old television, which can hide the very clots doctors are trying to find. Traditional methods to clean up this graininess often blur the fine details, making small blood vessels look fuzzy. Recently, a new type of computer processing called deep learning has emerged. Unlike older methods that simply smooth out the image, this technology learns from vast amounts of high-quality data to remove the grain while keeping the sharp edges of the anatomy intact.

A team of researchers at the Affiliated Hospital of Xuzhou Medical University set out to test whether combining these low-energy settings with this new deep learning technology could solve the problem. They wanted to see if they could create a scan that used significantly less radiation and less contrast dye, yet still provided a crystal-clear view of the smallest blood vessels in the lungs, where clots are most dangerous and hardest to see. They focused their study on patients suspected of having pulmonary embolism, scanning them with a dual-energy CT machine that can capture data at different energy levels simultaneously.

The researchers scanned 39 patients using a protocol designed to be gentle on the body, injecting a much smaller amount of dye than usual—about 26 milliliters on average—and keeping the radiation dose low. They then reconstructed the images using various computer settings. They tested different energy levels, ranging from very low to standard, and applied two different types of image cleaning: a traditional statistical method and the new deep learning method, which they tested at medium and high strengths. The goal was to find the perfect combination that would cancel out the noise of the low-energy scan without blurring the delicate details of the lung's peripheral arteries.

The results pointed to a clear winner. The best images came from using the lowest energy setting available, 40 kiloelectron volts, paired with the high-strength deep learning reconstruction. This specific combination produced images where the blood vessels were incredibly bright and distinct, while the background grain was almost completely gone. When the researchers measured the clarity of the images, this method outperformed the standard approach used in hospitals today. The standard method, which uses a higher energy level and traditional cleaning, struggled to show the smallest vessels with the same level of sharpness. In contrast, the new low-energy, deep-learning method made the tiny, subsegmental arteries look just as clear as the standard method, but with a much higher contrast between the blood and the surrounding tissue.

Two experienced radiologists reviewed the images without knowing which method was used to create them. They rated the quality of the new method as excellent, giving it the highest possible score. They found that the images were sharp enough to confidently identify clots even in the most distant branches of the lung's blood supply. Crucially, they found no difference in diagnostic quality between the new, low-dose method and the traditional, higher-dose method. This means that doctors could potentially switch to this new protocol and still make accurate diagnoses, but with a significant reduction in the burden placed on the patient.

The study also highlighted what did not work as well. When the researchers tried to use the low-energy setting with the traditional cleaning method, the images were too noisy to be useful, confirming that the old technology could not handle the extreme clarity of the low-energy scan. Similarly, using the deep learning method at a medium strength was not quite as effective as the high-strength version. The research suggests that the high-strength deep learning algorithm is essential for unlocking the benefits of low-energy scanning.

This work demonstrates that it is possible to drastically reduce the amount of radiation and contrast dye a patient receives without sacrificing the ability to see small, dangerous clots in the lungs. By finding the right balance between energy levels and advanced computer processing, the researchers have shown a path toward safer, more comfortable scans for patients. While the study was conducted at a single center and involved a specific type of scanner, the findings offer a promising blueprint for how hospitals might improve care in the future, ensuring that the tools used to save lives do not inadvertently cause harm through excessive exposure.