Explainable Hybrid ConvNeXt–Vision Transformer Model for Pneumonia Detection from Chest X-ray Images

Imagine a world where your eyes are the ultimate detectives, but sometimes, even the sharpest detective needs a little help seeing the invisible. In the realm of medicine, doctors use special cameras called X-rays to look inside our chests, trying to spot a sneaky invader called pneumonia. Pneumonia is a lung infection that makes breathing hard, and finding it early is like catching a fire before it burns the whole house down. But here's the tricky part: sometimes the difference between a healthy lung and an infected one is as subtle as a whisper in a storm. It's easy for even a human expert to miss the clues.

Enter the world of Artificial Intelligence (AI), specifically a branch called "Deep Learning." Think of Deep Learning as a super-smart student that learns by looking at thousands of pictures. Usually, this student uses two main ways to study: one way is like looking at a painting with a magnifying glass to see tiny brushstrokes (this is called a Convolutional Neural Network, or CNN), and the other is like stepping back to see the whole picture and how the colors relate to each other (this is called a Vision Transformer, or ViT). For a long time, scientists debated which student was better. But what if we could build a super-student that does both at the same time? That's exactly what a team of researchers from the International Islamic University Chittagong set out to do. They wanted to create a tool that doesn't just guess if a lung is sick, but also explains why it thinks so, helping doctors make faster, safer decisions.

The Super-Student's Secret Weapon

The researchers built a "hybrid" model, which is a fancy way of saying they glued two different AI brains together to make one super-brain. One half of this brain is based on ConvNeXt, which is excellent at spotting local details, like the tiny, fuzzy patches that often appear when a lung is infected. The other half is a Vision Transformer (ViT), which is great at understanding the big picture and how different parts of the lung connect.

But they didn't stop there. They added a special "attention mechanism" called CBAM. Imagine this as a pair of magical glasses that tell the AI, "Hey, look here, this part is important, ignore that blurry background." This helps the model focus on the exact spots in the X-ray that matter most, filtering out the noise.

To teach this new super-brain, the team didn't just use a few pictures. They gathered a massive library of 6,590 chest X-ray images. Some showed healthy lungs, and others showed pneumonia. They split this library into three piles: one for learning (training), one for checking homework (validation), and one for the final exam (testing). Before feeding these images to the AI, they played a few tricks to make the data even better, like flipping the images sideways or changing the brightness, so the AI wouldn't get confused if a picture looked slightly different.

The Results: A New Champion

When it came time for the final exam, the results were impressive. The hybrid model got it right 94.03% of the time. To put that in perspective, they tested it against some of the other top AI students in the room, like ResNet152, MobileNetV2, and even the Vision Transformer on its own. The hybrid model beat them all, scoring higher in accuracy, precision, and recall.

But the real magic wasn't just the score; it was the "explainability." In the past, AI models were like black boxes: you put an X-ray in, and a "Yes" or "No" popped out, but you had no idea why. The researchers used a technique called Grad-CAM to turn the AI's thinking process into a colorful heat map. When the model saw pneumonia, the heat map lit up the exact infected areas of the lung, glowing red to show the doctor, "I'm worried about this spot." This is a huge deal because it builds trust. A doctor can look at the X-ray, see the glowing red spot, and say, "Ah, I see what the computer sees," rather than blindly trusting a guess.

Why This Matters (and What It's Not)

This study suggests that combining different types of AI brains, along with a way to focus on important details, creates a more reliable tool for spotting pneumonia. The model showed it could distinguish between sick and healthy lungs with high confidence, achieving a precision of 95.01% and a recall of 92.74%. The researchers argue that this approach is better than using just one type of model because it captures both the tiny details and the big picture.

However, the paper is careful to note that this isn't a magic cure-all yet. The model still struggles a little when pneumonia looks very mild or when the images are tricky, and it was tested on a specific set of data. The authors suggest that while this is a promising step toward a computer-aided diagnostic tool, more work is needed to make it perfect for every hospital in the world. They also point out that the dataset, while large, could be even bigger and more diverse.

In short, this paper doesn't claim to have solved pneumonia forever. Instead, it offers a powerful, transparent new way to help doctors see the invisible, proving that when you combine the best of two different AI worlds and add a little bit of "attention," you get a tool that is not only smart but also easy to understand.