AppendiGrade: An XAI-Enhanced Deep Learning Framework for Grading Appendicitis in Ultrasound with Gaussian Blur and Grad-CAM

In the world of medicine, the sudden, sharp pain of appendicitis is a common emergency that demands quick action. If the inflamed appendix bursts, it can lead to life-threatening complications, making early and accurate diagnosis critical. While doctors have long relied on imaging tools to see inside the body, ultrasound stands out as a particularly valuable method because it uses sound waves rather than radiation, making it safer and more accessible for patients. However, looking at these images is not always straightforward. Distinguishing between a simple case of inflammation and a more dangerous situation, such as a burst appendix or a pocket of infection, requires a highly trained eye and can be time-consuming. This is where the intersection of medicine and computer science offers a new path forward, using artificial intelligence to help doctors see what might otherwise be missed.

A team of researchers set out to build a computer system capable of reading these ultrasound images and automatically identifying the specific type of appendicitis present. They focused on five distinct categories: a normal, healthy appendix; acute inflammation; a perforated appendix where a hole has formed; an abscess, which is a collection of pus; and an appendicolith, a hard, stone-like deposit inside the organ. To teach their computer, the team gathered a large collection of 4,679 ultrasound images, each carefully labeled by a certified physician to ensure the computer was learning from correct information. They began by feeding these raw images into four different pre-existing computer models, which are like digital brains designed to recognize patterns. Initially, the results were underwhelming. The models struggled to make sense of the images, achieving an accuracy of only about 69 percent, meaning they were wrong more than one-third of the time. The images contained too much visual noise and lacked the clear contrast needed for the computer to distinguish between the different types of inflammation.

Realizing that the raw data was the bottleneck, the researchers decided to clean up the images before showing them to the computer. They applied a process that first smoothed out the grainy noise, much like blurring a photograph to remove distracting specks, and then sharpened the edges to make the boundaries of the organs stand out clearly. They also adjusted the internal settings of the computer models, fine-tuning how much the system learned from each new example. This combination of cleaning the images and tuning the machine's settings transformed the results. The system that performed best, a model known as InceptionV3, saw its accuracy jump dramatically from 69 percent to 95.58 percent. It became highly reliable, correctly identifying the condition in almost every case it tested.

The researchers did not stop at just getting the right answer; they wanted to understand how the computer reached its conclusion. To do this, they used a technique that creates a visual map, highlighting the specific areas of the ultrasound image that the computer focused on to make its decision. When they looked at these maps, they saw that the computer was indeed looking at the correct parts of the body, such as the inflamed appendix or the surrounding fluid, rather than guessing based on random patterns. This transparency is crucial, as it allows a human doctor to quickly verify that the computer is looking at the right things. The study found that while other models performed well, this particular one was the most consistent and reliable, showing that with the right preparation of data, artificial intelligence can become a powerful tool for distinguishing between the subtle differences in abdominal emergencies. The work suggests that such systems could one day assist medical professionals by highlighting areas of concern on ultrasound scans, offering a second opinion that is fast, consistent, and grounded in the visual evidence of the image itself.