SCD-E2S: Deep Learning Based Skin Cancer Detection Using EfficientNetV2-S with Attention Mechanisms

Skin cancer is a growing global health challenge, with the most common forms appearing as dark spots, raised bumps, or irregular patches on the skin. Detecting these conditions early is vital, as it dramatically improves a patient's chances for a full recovery. For decades, doctors have relied on their eyes and experience to spot these warning signs, but human judgment can vary from one physician to another, and the visual differences between a harmless mole and a dangerous tumor can be subtle. In recent years, a branch of artificial intelligence known as deep learning has emerged as a powerful tool to assist in this task. These computer systems learn by studying thousands of medical images, gradually recognizing the complex patterns that distinguish healthy skin from disease. The goal is not to replace the doctor, but to provide a reliable second opinion that can catch what might otherwise be missed, especially in areas where specialist care is hard to reach.

A team of researchers has now developed a new system designed to make this technology faster, more accurate, and ready for real-world use. Their work focuses on a specific type of artificial intelligence architecture called EfficientNetV2-S, which acts as the brain of the system. To make this brain even sharper, the researchers added special attention mechanisms, which function like a spotlight that helps the computer focus on the most important details of a skin lesion while ignoring irrelevant background noise. They trained this system using a massive collection of over 14,000 dermoscopic images—highly magnified pictures of skin taken with a special camera. Before feeding these images to the computer, the team cleaned them up by removing distracting elements like hair or ruler marks and enhancing the contrast so that the edges of the spots were clearer. They also took care to balance the data, ensuring the computer saw enough examples of rare, dangerous cancers to learn how to spot them, rather than just the common, harmless ones.

The training process was a carefully staged effort. First, the computer learned the basics using the new images, and then it was allowed to fine-tune its understanding by looking at the deeper layers of its own knowledge. To prevent the system from memorizing the training images too perfectly—a mistake that would make it fail on new pictures—the researchers used a technique that mixed different images together during learning, forcing the system to understand the underlying concepts rather than just the specific pixels. They also adjusted the system's internal rules to pay extra attention to the most difficult cases, particularly the dangerous melanoma, ensuring it did not overlook them. When tested on images it had never seen before, including data from two other major medical databases, the system achieved an accuracy rate of over 97 percent. It correctly identified the three main types of skin lesions it was designed to find: melanoma, basal cell carcinoma, and melanocytic nevi.

What sets this work apart is not just the high score, but how the system was built to be useful outside of a laboratory. The researchers packaged their trained model into a mobile application that runs on standard smartphones. In this setup, a user can take a photo of a suspicious spot, and the image is sent securely to a server where the computer analyzes it in just over one second. The app then returns a result with a confidence score, acting as a decision-support tool for both patients and doctors. The researchers emphasize that this tool is designed to assist, not to replace, professional medical diagnosis. By combining a highly efficient computer model with a practical mobile interface, the team has demonstrated that advanced artificial intelligence can be made accessible for immediate use in clinics or remote areas. The study suggests that with further testing on diverse patient groups and integration into medical workflows, such systems could become a standard part of early skin cancer detection, helping to save lives by catching disease sooner.