TR-GS: High-Fidelity Sparse-View CT Volumetric Rendering via t-Distribution Gaussian Splatting and Ray-Confidence Modeling

Medical imaging has long relied on a delicate balance between clarity and safety. To see inside the human body without cutting it open, doctors use computed tomography, or CT, which builds a three-dimensional picture from hundreds of flat X-ray images taken from different angles. The more angles captured, the clearer the final picture, but each additional X-ray adds to the patient's radiation exposure. For decades, the goal has been to get a sharp, usable image from as few X-rays as possible, a challenge known as sparse-view reconstruction. When too few angles are used, the resulting images often suffer from streaks and blurring, making it difficult to spot small tumors or plan delicate surgeries. While modern computers can now generate 3D scenes from photos, applying these techniques to medical scans has been tricky because the math used for natural objects does not always handle the uncertainty and missing data found in low-dose medical scans.

A team of researchers at Shenzhen University has developed a new method called TR-GS to solve this specific problem. Instead of using the standard mathematical shapes that most 3D imaging software relies on, they replaced them with a different, more flexible shape that is naturally better at ignoring errors. In the world of computer graphics, 3D scenes are often built from millions of tiny, fuzzy ellipsoids, or oval-shaped clouds, that act like pixels in three dimensions. Standard methods use shapes that are very strict and thin-tailed, meaning they assume every piece of data is perfect and try to fit the shape exactly to it. When the data is incomplete or noisy, as it is in low-dose CT scans, these strict shapes struggle, leading to jagged artifacts and distorted structures. The researchers swapped these for a shape that has "heavier tails," a statistical property that allows the shape to be more forgiving of bad data points without losing the overall structure of the object.

To make this system work even better, the team added a smart layer of judgment to the process. They created a system that checks how much information each tiny 3D shape has received from the X-ray cameras. If a shape is surrounded by many X-ray beams coming from different directions, the system treats it as highly reliable and keeps its shape precise. If a shape is in a spot where very few X-rays pass through, the system recognizes the uncertainty and allows that shape to become more flexible and robust, preventing it from creating false details. This dynamic adjustment happens automatically during the computer's learning process, ensuring that well-observed areas remain sharp while poorly observed areas do not introduce confusing noise.

The researchers also introduced a technique to protect the fine details of the anatomy, such as the edges of bones or the texture of soft tissue, which can sometimes get smoothed out by the flexible shapes. They used a mathematical tool that separates the image into different layers of detail, allowing the computer to boost the sharp, high-frequency features in areas where the data is trustworthy, while suppressing random noise in areas where the data is weak. This ensures that the final image retains the crispness needed for medical diagnosis without amplifying the static that comes from using fewer X-rays.

When tested on both computer-generated models and real patient scans, this new approach produced significantly clearer images than previous methods. In tests using only 18 X-ray views, a very low number for a standard scan, the new method reconstructed the internal structures with a clarity score of 35.30, a metric that measures how close the image is to the perfect original. This was a noticeable improvement over the best existing techniques, which scored lower and showed more streaks and blurring. The method also proved more resilient when the X-ray data contained random noise, a common issue in low-dose scans, maintaining its accuracy where other methods began to falter.

The results suggest that this approach offers a viable path toward safer medical imaging. By allowing doctors to get high-quality 3D views from fewer X-ray shots, the technology could reduce the radiation patients receive during routine scans while still providing the detailed information needed for surgical planning and disease assessment. The researchers have made their code available to the public, inviting further development to apply these principles to even more complex medical scenarios, such as scans taken from limited angles or specific body regions where data is hardest to gather.