OPTIMIZING SKIN LESION CLASSIFICATION WITH CONFUSION-AWARE LOSS FUNCTIONS

Early diagnosis of skin cancer is critical to treatment and saving patients’ lives, many studies have used Convolutional Neural Networks (CNNs) to achieve this goal. Traditional methods using the Cross Entropy (CE) loss function, however, often struggle with classes that are easily confused, such as Nevus and Melanoma, leading to reduced diagnostic accuracy. To address this, we propose the Confusion-aware Cross Entropy (CCE) loss function, which enhances classification performance by focusing on these easily confused classes. Our method computes the mean of the negative class logits to identify these classes, ensuring the loss calculation prioritizes their accurate classification. Experiments conducted on the publicly available HAM10000 dataset using ResNet50, EfficientNet-B4, Inception-V3, and DenseNet121 demonstrate that our approach significantly outperforms the traditional CE loss function, achieving higher Accuracy, Sensitivity, and Precision. These results underscore the potential of the CCE loss function to improve clinical outcomes by providing more reliable skin lesion classifications.