AMPLIFAI: A Multiphase CT Dataset for Benchmarking Clinical Reasoning in LI-RADS Assessment of Liver Lesions
This paper introduces AMPLIFAI, the first publicly available dataset of multiphase abdominal CT scans annotated with LI-RADS categories and segmented key features, designed to overcome data scarcity and advance AI-driven, biopsy-free diagnosis of hepatocellular carcinoma.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Liver cancer is a silent and deadly adversary, ranking as the third leading cause of cancer-related death worldwide. The difference between life and death often comes down to timing: catching the disease early can raise a patient's chance of survival from less than one in five to more than seven in ten. For decades, doctors have relied on a strict set of rules called LI-RADS to decide if a spot on a liver scan is cancer. These rules do not rely on a single snapshot but on a movie-like sequence of images taken as a special dye moves through the body. By watching how a suspicious spot lights up and then fades away compared to healthy tissue, radiologists can determine if a biopsy—a painful needle test—is even necessary. While this system works well for human experts, teaching computers to see these same patterns has been nearly impossible because the right kind of training data simply did not exist.
This is where a new effort called AMPLIFAI steps in to fill a critical gap. Researchers at the University of Maryland have created the first public collection of liver scans specifically designed to teach artificial intelligence how to apply these complex rules. The dataset consists of 590 cases gathered from four different medical archives around the world. Each case is not just a single image but a full study containing multiple phases of a CT scan: a baseline image taken before any dye is injected, followed by a rapid sequence of images taken as the dye flows through the arteries, the veins, and finally lingers in the tissue. To make this data useful for machines, a team of five expert radiologists and a trainee spent approximately two months carefully labeling every scan. They did not just mark where a tumor was; they drew precise outlines around three specific behaviors that define the cancer rules: a bright flash of color in the early phase, a darkening effect as the dye leaves, and a thin, glowing rim that stays bright longer than the rest of the liver.
The result is a massive, organized library of 590 liver studies, representing 584 unique patients, that is now available for scientists to use. The researchers split this collection into a training group of 531 cases and a smaller testing group of 59 cases, ensuring that no patient appeared in both groups so that the computer models would be tested on truly new data. The annotations are incredibly detailed, providing a pixel-by-pixel map of the tumor and its specific features. This level of detail allows artificial intelligence to learn not just that a tumor exists, but exactly how it behaves over time. For instance, the data helps a computer understand that a tumor must show a specific type of bright flash that spreads from the inside out, rather than just a ring around the edge, to be considered a definite cancer. Without this specific distinction, a computer might misclassify a dangerous tumor as less severe, or vice versa.
The team behind this project, drawn from radiology and computer science departments, built a custom software tool to help the doctors draw these outlines efficiently. They used a method where the computer would make a first guess at the shape of the tumor, and the human experts would then refine it, a process that repeated as the computer learned from more examples. This ensured that the final labels were consistent and accurate, reducing the natural variations that can happen when different doctors look at the same image. The dataset includes cases ranging from clearly benign spots to confirmed cancers, as well as rare cases where cancer has invaded blood vessels or where the tumor looks malignant but does not fit the typical pattern of liver cancer.
By releasing this data under a license that allows for non-commercial research, the authors hope to accelerate the development of tools that can assist doctors in diagnosing liver cancer earlier and more reliably. The dataset is not a finished medical device; it is a foundation. It provides the raw material needed to train the next generation of AI systems to recognize the subtle, shifting patterns of disease that human eyes have long used to save lives. While the data itself cannot diagnose a patient, it offers a clear path for researchers to build systems that might one day help ensure that no case of liver cancer goes undetected until it is too late.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.