BoneNet: A Multimodal Deep Learning Framework with Dual-Modality Explainability for Automated Peri-Implant Bone Loss Detection and Severity Grading

Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints, you are looking for tiny cracks in a tooth's foundation. In the world of dentistry, when a metal screw (an implant) is placed into the jawbone to hold a fake tooth, the bone around it can sometimes start to shrink away. This is called "peri-implant bone loss." If a dentist misses this shrinking, the fake tooth can fall out, leading to painful and expensive repairs. Usually, dentists have to squint at X-ray pictures and use rulers to measure these tiny gaps by hand. It's slow, and different dentists might measure the same gap differently.

Enter the world of Artificial Intelligence (AI), specifically a type called "Deep Learning." Think of this as teaching a computer to see patterns the human eye might miss, like a super-powered student who has studied millions of X-rays. The goal is to build a computer program that can not only spot the shrinking bone but also tell you exactly how bad it is and how deep the gap goes. But here's the tricky part: computers are great at spotting where something is, but they sometimes struggle to understand how serious it is, especially when the clues are subtle. This paper is about building a smarter detective that doesn't just look at the picture but also does the math to get the answer right.

---

The Paper: BoneNet, the Super-Detective Dentist

The researchers behind this study, Muralidhar Billa and his team from RV University, wanted to build the ultimate AI tool for spotting bone loss around dental implants. They didn't just want a tool that says, "Hey, there's a problem here!" They wanted one that could grade the problem from "No Worries" to "Critical Emergency" and even measure the depth of the bone loss in millimeters.

The Detective's Training: Trying Out Nine Outfits
Before building their own special tool, the team decided to test nine different existing AI "outfits" (architectures) to see which one was the best at finding bone loss. They treated these like different pairs of shoes, trying them all on a balanced set of 519 X-ray images. Most of the shoes fit well, but one stood out: a model called YOLOv8m. It was the fastest and most accurate at finding the bone loss, but the team felt it could be improved. They realized that bone loss isn't just one size; it has tiny, fine textures and larger, broader areas. The standard YOLOv8m was a bit like wearing one-size-fits-all shoes—it worked, but it wasn't perfectly tailored to the specific shape of the problem.

The Upgrade: BoneNet
So, the team built their own custom detective, which they named BoneNet. They took the best parts of the YOLOv8m and swapped out its internal engine. Instead of standard blocks, they installed something called Dilated Residual Attention Blocks (DRAB).

• The Analogy: Imagine looking at a map. A normal camera zooms in on one spot. The DRAB is like having two pairs of glasses at the same time: one with a wide-angle lens to see the big picture of the bone, and another with a zoom lens to see the tiny, fuzzy textures of the bone cracks. By combining these views, BoneNet can see the bone loss much more clearly.
• The Result: This upgrade made BoneNet incredibly accurate. It achieved a score of 0.9660 (a measure of how well it finds and outlines the bone loss), which was a 3.83-point improvement over the best standard model. Plus, it was actually lighter and faster, using fewer computer resources.

The Secret Weapon: Adding a "Math Brain"
Here is where the paper gets really interesting. The team realized that even with a super-accurate detector, the AI was still struggling to guess the severity of the bone loss. It could find the spot, but it wasn't sure if it was a Grade 1 (mild) or Grade 4 (severe) problem.

• The Problem: Looking at the picture alone wasn't enough. It's like trying to guess how heavy a suitcase is just by looking at a photo of it; you can't tell if it's full of feathers or bricks.
• The Solution: They gave the AI a "Math Brain." They took the geometric measurements that a human expert would calculate from the X-ray (like the area, perimeter, and shape of the bone loss) and fed them into the AI as numbers.
• The Fusion: They used a special "cross-attention" network to let the picture and the numbers talk to each other.
• The Result: This was a massive game-changer. The AI's ability to guess the severity grade jumped from 50.79% (basically guessing) to 85.96% when it used both the picture and the math. This proves that for this specific task, you can't just rely on the image; you need the geometric numbers too.

Measuring the Depth: The Ruler Test
The team also wanted the AI to measure exactly how deep the bone loss was, in millimeters. They trained a separate math-focused AI (called XGBoost) to do this.

• The Result: On average, the AI was off by only 0.638 mm. In the world of dentistry, being within ±1 mm is considered acceptable, so this is a success!
• The Catch: The paper is very honest about a flaw. While the average was good, the AI got much worse as the bone loss got worse. For mild cases, it was perfect. But for the most severe cases (Grade 4), the error grew, and only 42.9% of those predictions were within the safe 1 mm range. The authors explain this is because they didn't have enough examples of severe cases in their training data to teach the AI how to handle the worst scenarios.

A Critical Note on the Detective's Training Data
There is one important detail about how the team trained and tested their detective that you should know. The study had 135 X-ray images, but some of these images contained multiple implants. When they split the data to test the AI, they separated the implants rather than the patients. This means that parts of the same X-ray image could end up in both the training group and the testing group.

• Why this matters: It's like giving the detective a practice test that includes questions they've already seen in their study guide because they came from the same source. This creates a "data leakage" risk, meaning the test scores might be slightly higher than what the AI would achieve if it saw completely new patients in the real world. The authors are transparent about this, noting that their reported accuracy is likely an "upper bound" (a best-case scenario) until they can test the system with data split strictly by patient.

The "Why" and "How": Making it Transparent
Finally, the team wanted to make sure the AI wasn't just guessing randomly. They used two tools to peek inside the AI's brain:

1. EigenCAM: This highlights the parts of the X-ray the AI is looking at. They found that for mild cases, the AI looked at the top of the bone, but for severe cases, it looked deep down the side of the implant. This matched what human dentists do.
2. SHAP: This tool looked at the numbers. It showed that the AI was paying attention to the "perimeter" (the length of the edge) of the bone loss, which makes perfect sense: a longer, more jagged edge usually means a deeper problem.

The Bottom Line
The paper concludes that BoneNet is a highly effective tool for spotting and grading bone loss, but it works best when it combines the "eyes" of the image with the "math" of geometric measurements. However, the authors are careful to say that while the tool is great for mild to moderate cases, it still needs more training on severe cases before it can be fully trusted to measure the deepest bone loss on its own. They also note that because of the way the data was split, the current results might be slightly optimistic. They suggest that for now, it's a fantastic assistant for dentists to help them triage patients, but a human should always double-check the most severe cases.