ATLAS: Development and Evaluation of an AI -Enabled Framework for Anatomically Traceable Measurement Assurance in Fetal-Head Ultrasound Biometry

In the quiet, dimly lit room of an ultrasound clinic, a sonographer guides a probe over a pregnant woman's abdomen, searching for a specific view of the developing baby's head. This moment is critical. To understand how the baby is growing, doctors need precise measurements of the skull, specifically the width across the brain and the distance from front to back. These numbers help determine the baby's age and whether it is growing at a healthy pace. However, getting these numbers right depends entirely on the quality of the image. If the ultrasound probe is tilted even slightly, or if the baby is in an awkward position, the resulting picture might look like a head, but the measurements taken from it will be wrong. For decades, this has been a human challenge: a doctor must look at a grainy, black-and-white image, decide if it is the right angle, and then place digital calipers to measure the skull. As artificial intelligence begins to take over this task, promising to measure the baby's head automatically, a new problem has emerged. A computer can confidently output a number even if the image it is looking at is blurry, tilted, or missing key features. The danger is that a doctor might trust a perfect-looking number that was calculated from a flawed picture, leading to incorrect medical decisions.

A team of researchers at Griffith University and Gold Coast University Hospital has developed a new system called ATLAS to solve this problem of trust. Instead of simply giving a single number, ATLAS is designed to act like a careful assistant that checks its own work before speaking. The system breaks the measurement process into three distinct steps that mirror how a human doctor thinks: first, it finds the baby's head; second, it confirms that the internal structures of the brain are visible and correct; and third, it calculates the measurements. Crucially, the system keeps a detailed record of every step it took. It remembers exactly where it found the head, what evidence it saw to prove the image was valid, and how it converted the pixels on the screen into millimeters. This approach ensures that if a measurement is made, there is a clear, auditable trail of evidence showing why that measurement is reliable. The researchers tested this system on thousands of ultrasound images from different hospitals and found that while the computer could find the head almost perfectly, the way it measured the head mattered greatly. They discovered that a simpler method of measuring, which relied on the shape of the box the computer drew around the head, was often more accurate than a more complex method that tried to trace the exact outline of the skull. Furthermore, they showed that when the system was adapted to work with data from a new hospital, it could get better at finding the head but worse at measuring it, proving that finding an object is not the same as measuring it correctly.

The core of the ATLAS system is built on the idea that a measurement is only as good as the conditions under which it was made. In the first stage, called "Find," the computer scans the entire ultrasound image to locate the fetal head. It draws a box around the region it believes contains the skull. If the computer cannot find a head with enough confidence, it simply stops and reports that no measurement is available, rather than guessing. In the second stage, "Confirm," the system zooms in on that box to look for specific internal landmarks, such as a fluid-filled space in the middle of the brain and the lateral ventricles. These structures are essential for confirming that the ultrasound plane is the correct one for measurement. If these landmarks are missing, the system knows the image might be from the wrong angle, even if the head itself was found. Finally, in the "Measure" stage, the system calculates the width and circumference of the head. The researchers tested two different ways to do this: one method used a complex algorithm to trace the exact curve of the skull, while the other used the simple dimensions of the box found in the first stage.

When the researchers tested these methods on a large set of images from a public hospital, they found a surprising result. The simpler method, which used the box dimensions, produced more accurate measurements than the complex tracing method. Specifically, the box method was correct within a very tight margin for over 91 percent of the images, whereas the complex tracing method was correct for only about 84 percent. The complex method tended to underestimate the width of the head by nearly a millimeter on average, a small difference that could still shift a baby's growth chart significantly. This finding challenged the common assumption in artificial intelligence that more complex models are always better. The researchers also tested the system on images from two other hospitals in Spain and the United Kingdom to see if it would work on data it had never seen before. The system was excellent at finding the head in these new images, locating it in nearly every single frame. However, the accuracy of the measurements varied depending on the hospital. On the Spanish data, the simple box method remained the most accurate. On the UK data, the complex tracing method performed better. This showed that the best way to measure depends on the specific characteristics of the images being used, such as how the ultrasound machine was set up and how the images were labeled by experts.

A significant part of the study focused on what happens when a system is trained on data from one hospital and then used in another. The researchers tried to "fine-tune" the system using data from the new hospitals, hoping to make it work better there. They found that while this process made the system better at finding the head—improving the overlap between the computer's box and the actual head by a few percent—it actually made the measurements less accurate. The measurements became more wrong, not less. This suggests that simply making a computer better at spotting an object does not guarantee it will measure that object correctly. To solve this, the researchers tried a technique called aggregation, where they combined the knowledge from the different hospital-specific versions of the system without mixing the actual patient images. This approach partially restored the accuracy, particularly for the UK data, showing that it is possible to share learning between institutions without sharing private patient data. The entire system is remarkably lightweight, requiring only 16.5 megabytes of memory and taking less than 60 milliseconds to process a single image, which means it could potentially run on standard ultrasound machines without needing powerful, expensive computers.

The study also included a review by a medical specialist who looked at the images the system processed. The doctor confirmed that the system's step-by-step approach was easy to understand and that it correctly identified when an image was suitable for measurement and when it was not. The doctor noted that even when an image was not perfect for a standard measurement, the system still preserved useful information about the baby's position and anatomy. This is important because it means the system does not just throw away "bad" images; it keeps the evidence so a human can decide if there is still value in looking at them. The researchers emphasized that their system is not a replacement for a doctor's judgment but a tool to make the measurement process transparent. By keeping a record of every step, from the initial search for the head to the final calculation, the system allows doctors to see exactly why a number was generated. If a measurement seems off, the doctor can trace it back to see if the head was found in the right place, if the internal brain structures were visible, and if the correct scale was used to convert the image into millimeters.

The implications of this work extend beyond just measuring a baby's head. It addresses a fundamental issue in medical artificial intelligence: the need for systems that are not only accurate but also trustworthy and explainable. In the past, AI systems often acted as "black boxes," providing an answer without explaining how they reached it. This made it difficult for doctors to know when to trust the machine and when to double-check the work. ATLAS demonstrates that it is possible to build a system that treats measurement as a process of evidence gathering rather than just a calculation. The study showed that data integrity is just as important as the algorithm itself; the researchers built a layer of checks to ensure that the data used to train the system was clean and that no patient's data was accidentally mixed between training and testing groups. They even tested the system by deliberately introducing errors into the data, such as duplicate images or missing labels, and the system successfully caught every single one. This level of rigor ensures that the results are reliable and that the system is not learning from mistakes in the data.

Ultimately, the ATLAS framework offers a new way to think about medical AI. It shifts the focus from simply getting a number to understanding the conditions under which that number is valid. The researchers found that the system could handle the complexity of real-world ultrasound images, working across different hospitals and with different types of equipment. While the system is not yet ready for use in a live clinical setting, the study provides a strong foundation for future development. It shows that by breaking down a complex task into smaller, verifiable steps and keeping a clear record of the evidence, we can build AI systems that doctors can trust. The next step will be to test this system in real-time during actual ultrasound scans to see how it performs when a doctor is holding the probe and making decisions in real time. Until then, the work stands as a proof of concept that transparency and traceability are essential for the safe and effective use of artificial intelligence in medicine.