AI Model Shows Promise in Reducing Skin Cancer Detection Bias

A new artificial intelligence (AI) model designed to detect skin cancer demonstrated significant improvements in addressing potential biases related to skin tone and excessive hair, according to research published in BMC Medical Informatics and Decision Making.

The model, called SkinWiseNet (SWNet), achieved greater than 99% accuracy in classifying skin lesions as benign or malignant by combining insights from different datasets incorporating “feature fusion,” a technique that combines insights from multiple datasets to reduce bias present in any singular dataset. The researchers, led by Ali Atshan Abdulredah of the National School of Electronics and Telecoms of Sfax, University of Sfax, Tunisia, found this approach particularly helpful for analyzing images that have historically posed challenges for AI detection systems.

“The proposed model addresses potential biases associated with skin conditions, particularly in individuals with darker skin tones or excessive hair,” the study authors wrote. “SWNet’s ability to classify normal and abnormal classes and its integration of feature fusion to mitigate biases reinforce its robustness and reliability in addressing diverse skin conditions.”

While artificial intelligence has transformed skin cancer detection over the past two decades, with studies showing a 15% to 20% improvement in prediction accuracy, most current AI systems still face challenges. Traditional diagnosis relies on visual examination, dermoscopy and biopsy, with AI systems now helping to automate and enhance this process by analyzing images layer by layer, similar to how human vision works. However, many existing AI models struggle with analyzing darker skin tones or images where hair is present.

The researchers addressed this issue by combining insights from four major skin cancer image databases. This approach allowed SWNet to learn from a more diverse set of examples than previous AI systems, which often trained on more limited datasets. As a result, SWNet’s 99.86% accuracy significantly outperforms other leading AI systems, including EfficientNet (91.88%), MobileNet (93.2%), and Darknet (94.44%).

Beyond achieving higher accuracy, SWNet demonstrated superior ability to explain its decision-making process, helping clinicians understand why the model classified an image as benign or malignant. The results showed this approach was particularly effective in avoiding common AI pitfalls, such as misinterpreting microscope artifacts or struggling with non-circular lesion patterns that often appear in real-world conditions.

The research team is currently working on addressing several limitations, including the challenges of handling large sets of high-quality images and optimizing the model's parameters for clinical use.