VIT-DRN-OCAFF: An Optimized Cross-Attention Feature Fusion Framework for Multi- Class COVID-19 and Pulmonary Disease Diagnosis Using Thoracic CT Images

The human lung is a complex, delicate organ, and when it falls ill, the signs can be subtle and easily confused. A cough, a fever, or a feeling of breathlessness might signal a common viral infection, a bacterial pneumonia, or a more severe condition like COVID-19. For doctors, distinguishing between these causes is critical, yet the symptoms often overlap, making a quick and accurate diagnosis a significant challenge. To see inside the chest without surgery, medical professionals rely on computed tomography, or CT scans, which create detailed cross-sectional images of the lungs. These images reveal the texture of the lung tissue, showing areas of infection, fluid, or inflammation. However, reading thousands of these scans is a heavy burden for radiologists, and the visual differences between similar diseases can be so slight that even experts might miss a crucial detail or misidentify the cause.

In recent years, computers have begun to assist in this task, using artificial intelligence to learn what different diseases look like on a scan. Early attempts used standard image-recognition tools, but these often struggled to see the big picture of how a disease spreads through the lung while also noticing the tiny, specific spots where the infection is most active. A new study proposes a more advanced way to teach a computer to read these scans, combining two different ways of looking at the image to create a clearer picture of the disease. The researchers aimed to build a system that could not only tell if a patient is sick but also pinpoint exactly what kind of illness they have, whether it is COVID-19, viral pneumonia, bacterial pneumonia, or other non-viral lung issues.

The researchers developed a new computer framework designed to act like a pair of expert eyes working together. One part of the system is trained to look at the entire lung image at once, understanding the broad patterns of how an infection might spread across the organ. This part is good at seeing the "forest," recognizing the general shape and distribution of the disease. The second part of the system focuses intensely on the small details, zooming in to find the specific textures and edges of the infected areas, like the "trees" in that forest. By itself, the first part might miss a small but critical spot of infection, and the second part might get lost in the details without understanding the overall context. The innovation in this study lies in how these two parts are connected. The researchers created a special mechanism that allows the two systems to talk to each other, constantly checking and refining what they see. This process ensures that the computer prioritizes the most important features for diagnosis while ignoring irrelevant noise or confusing patterns that do not help identify the disease.

To test this new approach, the team gathered a massive collection of lung scans, totaling nearly 25,000 images. These images came from public medical databases and from hospitals in India, covering a wide range of patients and disease severities. The dataset included scans from people with confirmed COVID-19, viral pneumonia, bacterial pneumonia, and those with other lung abnormalities that were not caused by a virus. Before the computer could learn from these images, the researchers carefully cleaned and prepared them. They removed unnecessary parts of the body from the scans, adjusted the brightness and contrast to make the infected areas stand out, and smoothed out any grainy noise that might confuse the system. This preparation ensured that the computer was learning from the actual signs of disease rather than artifacts caused by the scanning machine.

The computer was then trained on a large portion of these images, learning to recognize the unique signatures of each disease. Once the training was complete, the system was tested on a separate set of images it had never seen before to see how well it could perform in a real-world scenario. The results were remarkably high. The system correctly identified the type of lung disease with an AUC of 0.99. It was particularly good at avoiding false alarms, correctly identifying healthy lungs or non-disease conditions when they were present, and it rarely missed a true case of infection. When the researchers broke down the performance by disease type, the system excelled at distinguishing COVID-19 from other infections, achieving nearly perfect accuracy. It also performed strongly in separating viral pneumonia from bacterial pneumonia, two conditions that are often difficult to tell apart even for human doctors.

To understand why this new system worked so well, the researchers compared it against older methods. They tested standard computer vision tools, traditional deep learning models, and even other advanced systems that used only one of the two approaches described above. The older methods, while better than random guessing, consistently fell short, with accuracy rates ranging from the mid-80s to the high-90s percent. The new system, which combined the broad view and the detailed view with its special communication link, outperformed all of them. The study also showed that the system was efficient, processing images quickly enough to be useful in a busy hospital setting without requiring excessive computing power.

The researchers did not stop at just numbers; they also looked at how the system "saw" the images. They compared the areas the computer highlighted as infected against the actual infected areas marked by human doctors. The computer's focus matched the human experts' findings with high precision, correctly identifying the boundaries of the infection and the specific textures associated with each disease. This level of agreement suggests that the system is not just guessing based on patterns but is actually learning the visual characteristics of the diseases. The study concludes that this new framework offers a highly reliable tool for diagnosing lung diseases from CT scans. By combining a broad understanding of the lung's condition with a sharp focus on specific lesions, the system provides a level of accuracy that could significantly aid doctors in making faster and more confident decisions for their patients. While the study was conducted on a specific set of data, the results suggest a powerful step forward in using artificial intelligence to support the diagnosis of respiratory illnesses.