PROJECT TITLE :

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

ABSTRACT:

Melanoma is a form of skin cancer that is curable and has a dramatically increasing survival rate when diagnosed at an early stage. However, it is also a cancer that almost always results in death. The detection of melanomas in dermoscopic images using learning-based methods holds a great deal of promise. However, because melanoma is such a rare disease, most existing databases of skin lesions contain a highly imbalanced number of benign samples compared to malignant ones. As a result, this imbalance causes significant bias to be introduced into classification models as a consequence of the majority class's preponderance in statistical analysis. In order to solve this problem, we have devised a method of deep clustering that is predicated on the latent-space embedding of dermoscopic images. Clustering is accomplished by enforcing a novel center-oriented margin-free triplet loss (COM-Triplet) on image embeddings produced by a convolutional neural network backbone. This creates the necessary conditions for the clustering process. The objective of the proposed method is not to achieve the lowest possible classification error; rather, it seeks to create cluster centers that are maximally distinct from one another. As a result, it is less sensitive to class imbalance. We further propose to implement COM-Triplet based on pseudo-labels generated by a Gaussian mixture model in order to circumvent the requirement for labeled data. This will allow us to avoid the burden of collecting labeled data (GMM). Extensive testing has shown that deep clustering with COM-Triplet loss performs better than traditional clustering with triplet loss as well as other competing classifiers in both supervised and unsupervised environments.