Enhancing Clinical Decision Support Through Explainable AI: A Comparative Study of CNN and Vision Transformer Models for Skull Fracture Detection in CT Scans

Imagine you are a detective trying to solve a mystery inside a locked room. In the world of medicine, that room is the human head, and the mystery is a hidden crack in the skull—a fracture caused by a bump or a fall. To see inside without breaking the door, doctors use a special kind of X-ray called a CT scan, which takes hundreds of thin slices of the brain like pages in a book. But reading all those pages is exhausting work. Sometimes, the cracks are so thin and tricky that they look just like normal lines in the bone, and even expert doctors can miss them. This is where "Artificial Intelligence" (AI) steps in. Think of AI as a super-fast robot assistant that can read all the pages in seconds. But here's the catch: robots are often like black boxes. They give you an answer, but they won't tell you why they think that. If a robot says, "There's a crack here," but you can't see where it's looking, a doctor might not trust it. This paper is about building a robot that not only finds the cracks but also points its finger at exactly where it saw them, so doctors can say, "Ah, I see what you mean."

The researchers in this study decided to put two different types of robot detectives to the test: one that thinks like a traditional human eye (called a CNN, specifically ResNet-50) and one that thinks like a modern, pattern-hunting super-organism (called a Vision Transformer, specifically DeiT-Tiny). They wanted to see which one was better at spotting skull fractures in CT scans and, more importantly, which one could explain its thinking clearly enough to be trusted in a hospital.

The Great Detective Showdown

To run this experiment, the team gathered a collection of 4,914 CT scan slices from 15 patients who had suffered head injuries at a hospital in Sri Lanka. It's a bit like having a small library of mystery books rather than a massive one, which makes the job harder for the robots. They cleaned up the images, making the bones look sharp and clear, and then taught the two robots how to spot fractures.

The first robot, ResNet-50, is like a detective who zooms in on tiny details. It looks at small patches of the image, checking every little bump and line. The second robot, DeiT-Tiny, is more like a detective who steps back to look at the whole picture at once, connecting dots across the entire image to find patterns.

When the robots started their work, the results were clear. The traditional detective, ResNet-50, was the better finder. It correctly identified fractures about 72.6% of the time and had a "score" for how good it was at distinguishing cracks from normal bone (called AUC-ROC) of 0.800. The modern detective, DeiT-Tiny, was a bit slower and less accurate, getting the right answer 67.3% of the time with a score of 0.746. The difference wasn't just a fluke; the math showed it was a real, significant gap. In fact, ResNet-50 was much better at avoiding false alarms (saying there was a crack when there wasn't one), while DeiT-Tiny tended to get a little too excited and see cracks where there were none.

The "Why" Behind the Answer

But being a good detective isn't just about finding the answer; it's about showing your work. This is where the story gets interesting. The researchers wanted to know: When the robot says "fracture," where is it actually looking?

To check this, they asked a human expert to draw boxes around the real fractures on a few test images. Then, they compared those boxes to the "heatmaps" the robots drew to show their focus. Imagine the robots drawing a glowing cloud over the image to show where they think the crack is.

Here, the tables turned slightly. While ResNet-50 was better at finding the cracks, DeiT-Tiny was slightly better at pointing to them. The modern robot's glowing cloud overlapped with the real fracture a bit more than the traditional robot's did. However, the overlap was still quite small for both—think of it as the robot pointing generally in the right neighborhood, but not quite at the exact house number. The numbers were low: DeiT-Tiny managed a tiny overlap score of 0.021, while ResNet-50 was at 0.006. Even the "better" robot wasn't perfect at pinpointing the exact spot.

What This Means for the Future

So, what's the verdict? The study suggests that if you need a robot to tell you if a patient has a fracture, the traditional ResNet-50 is currently the more reliable choice, especially when you don't have a huge amount of data to train on. It's the steady, experienced detective who gets the job done. However, if you care about the robot being able to show you where it's looking in a way that feels a bit more logical to a human eye, the modern DeiT-Tiny has a slight edge, even if it's not as good at the actual detection.

The researchers are careful to say that neither robot is ready to replace a doctor just yet. Both robots missed about half of the actual fractures (a "recall" of around 0.49), meaning they still need a human to double-check their work. But the study highlights a fascinating trade-off: sometimes the robot that is best at solving the mystery isn't the one that explains its reasoning the best.

The team concludes that for these robots to become true partners in the hospital, we need to teach them on more data from different hospitals and help them get even better at pointing out the exact location of the injury. Until then, they serve best as a helpful second pair of eyes, ready to flag suspicious slices for the radiologist to take a closer look, ensuring that no hidden crack goes unnoticed.