BoneAgeTW2: Automated Skeletal Maturation Assessment via the Tanner-Whitehouse 2 Method, Deep Learning, and Clinical Report Generation with Distribution Curves

Imagine your body is a construction site, and your bones are the steel beams being erected. While your birthday tells you how many years you've been on the job site, your bones have their own internal clock, growing and hardening at their own pace. Sometimes, this biological clock runs fast, sometimes slow, and doctors need to know the difference to help kids with growth problems or to understand how much time is left for a teenager to grow taller. For decades, the "gold standard" for checking this bone clock has been a method called Tanner-Whitehouse 2 (TW2). It's like a master builder inspecting 20 specific steel beams in the hand and wrist, checking each one to see if it's still soft and growing or if it's fully hardened. This inspection is incredibly detailed but also incredibly slow, taking a human expert 10 to 20 minutes per patient, and even then, two experts might disagree by up to a year.

Now, enter a new digital tool designed to speed up this process without losing the detail. This paper introduces BoneAgeTW2, an artificial intelligence system that acts like a super-fast, tireless apprentice radiologist. Instead of just guessing a single number for a child's bone age, this system mimics the human expert's exact process. It looks at an X-ray of a hand, finds 20 specific bones, and assigns a maturity stage to each one, just like a human would. It then crunches the numbers to give a final age estimate, complete with a confidence range and a visual report showing exactly how each bone compares to the average for that age. The goal isn't just to be fast; it's to be transparent, giving doctors a clear, step-by-step breakdown of why the system reached its conclusion, rather than just a mysterious black-box answer.

The Problem: The Slow, Human Way

For a long time, checking a child's bone age has been a bit like reading a very complex, handwritten map. The Tanner-Whitehouse 2 (TW2) method is the map everyone in Latin America and Europe uses. To read it, a radiologist has to look at an X-ray of a hand and identify 20 specific bones, from the long ones in the fingers to the small, puzzle-piece bones in the wrist. For each bone, they have to decide which "stage" of maturity it is in, ranging from a soft, early stage (A) to a fully hardened, adult stage (I).

Once they've scored all 20 bones, they have to add up the points using special tables to get a final "bone age." To make sense of the result, they often draw bell-shaped curves (Gaussian distributions) to see if a specific bone is ahead of schedule or lagging behind. The problem? This whole process takes 10 to 20 minutes per patient. It requires years of training, and because it's so manual, two different doctors might give different answers for the same X-ray, sometimes by as much as 12 months. While there are some commercial tools that can guess a bone age, they are expensive, closed-source (you can't see how they work), and they usually just spit out a single number without showing the work behind it.

The Solution: BoneAgeTW2

The authors of this paper built BoneAgeTW2, an open-source system that automates the entire TW2 workflow. Think of it as a digital assembly line that takes an X-ray in one end and spits out a full clinical report in under 5 seconds.

Here is how the machine works, step-by-step:

1. The Setup: First, the system cleans up the X-ray image. It checks the digital file to see if the image is magnified (like a photo taken too close to a window) and corrects the size automatically. It also flips the image if it's a right hand, ensuring the computer always looks at a "left hand" view, just like its training data.
2. Finding the Bones: Using a smart detector called YOLOv8, the system hunts down the 20 specific bones in the hand. It draws a box around each one, acting like a highlighter pen marking the spots the radiologist needs to check.
3. The Classification: This is the tricky part. The system needs to know which stage (A through I) each bone is in. Since the training data didn't have labels for every single bone's stage, the authors used a clever trick called "pseudo-labeling." They used the known bone age of the patient to guess the most likely stage for each bone, creating a training set. Then, they trained a deep learning model (EfficientNet-B3) with 20 different "heads," where each head is an expert on just one specific bone.
4. The Scoring: Once the system has a stage for every bone, it adds up the points exactly according to the official TW2 rules. It calculates a score for the long bones (RUS) and the wrist bones (Carpal).
5. The Report: Finally, it generates a full report. This isn't just a number. It includes an annotated image showing the 20 boxes, a table with the score for every single bone, and—this is the cool part—interactive bell curves. These curves show the doctor exactly where the patient's bone sits compared to the average population. If a bone is delayed, the doctor can see the curve peaking far to the left of the patient's actual age.

What the System Achieves (and What It Doesn't)

The paper presents BoneAgeTW2 as a system that successfully automates the entire TW2 process, from raw image to a downloadable PDF report. The system is designed to run on standard computer hardware (CPU) in under 5 seconds, or even faster on a graphics card (GPU).

The authors are careful to manage expectations regarding accuracy. They acknowledge that because the system is forced to follow the rigid, step-by-step rules of the TW2 method (which uses discrete stages like A, B, C), it might not be as mathematically precise as some "black box" AI systems that just guess a number directly. In fact, they expect their system to have a slightly higher margin of error (Mean Absolute Error) than those simpler AI models. However, they argue this is a good trade-off. The goal isn't just to be the most accurate number-cruncher; it's to be explainable.

By breaking the answer down bone-by-bone, the system allows a doctor to see which bones are causing a delay or an advance. If a child's bone age is low, the doctor can look at the report and see, "Ah, the radius is at stage F, but the ulna is at stage E," giving them a much clearer picture of the patient's condition than a single number ever could.

The "Bell Curve" Feature

One of the most unique features of BoneAgeTW2 is how it visualizes the data. For every one of the 20 bones, the system draws a bell curve. Imagine a graph where the horizontal line is age, and the curve shows how common a certain bone stage is at different ages.

• If the patient's bone is at stage "E," the system highlights the "E" curve.
• It then draws a vertical line for the patient's actual age.
• If the line is far to the right of the curve's peak, the doctor knows that bone is "delayed" (it's acting younger than the patient is).
• If the line is to the left, the bone is "advanced."

This visual tool, which is usually drawn by hand or skipped entirely due to time constraints, is generated automatically for every bone, giving a rich, detailed view of the child's skeletal maturity.

The Bottom Line

BoneAgeTW2 is a proof-of-concept system that brings the detailed, human-like reasoning of the Tanner-Whitehouse 2 method into the digital age. It doesn't replace the radiologist; instead, it acts as a powerful assistant that does the heavy lifting of finding bones, scoring them, and drawing the charts in seconds. The authors emphasize that this is a decision-support tool that must still be reviewed by a qualified doctor, especially since the system relies on training data that might not perfectly match every population (like children in Colombia, where the study originated).

While the system is not yet a "solved" medical device ready for every hospital, it demonstrates that it is possible to automate a complex, manual medical protocol while keeping the logic transparent. It turns a 20-minute, expert-only task into a 5-second process that produces a report so detailed, a doctor can see exactly how the AI reached its conclusion, bone by bone.