Deep Clustering via Center-Oriented Margin-Free Triplet Loss for Skin Lesion Detection in Highly Imbalanced Datasets

The study addressed the challenge of detecting melanoma, a rare but fatal skin cancer, in highly imbalanced datasets of dermoscopic images, where benign samples vastly outnumber malignant ones. To mitigate the bias introduced by this imbalance, the researchers developed a deep clustering method using a novel center-oriented margin-free triplet loss (COM-Triplet) applied to image embeddings from a convolutional neural network. This approach focused on creating well-separated cluster centers rather than minimizing classification error, making it less sensitive to class imbalance. Additionally, the method utilized pseudo-labels from a Gaussian mixture model to eliminate the need for labeled data. Experiments demonstrated that the COM-Triplet loss method outperformed traditional triplet loss and other classifiers in both supervised and unsupervised scenarios.