Authors - Rajendra Yadav, Pavan Kumar Boori, Gaurav Kumawat, Chirag Joshi Abstract - Examining skin images from dermatoscopes can be a challenging endeavor especially when unlabelled skin images are being used. Therefore, we propose a novel method based on deep learning called Contrastive Clustering Autoencoder (CCA), specifically for clustering skin lesions without using la-belled data. CCA employs contrastive learning as well as clean autoencoder architecture based off a truncated ResNet-18 architecture. Internally, these models separate features into two branches: one for reconstruction of the image and one for clustering similar lesions, allowing it to learn informative patterns while maintaining image quality. Contrastive loss is used to create tight clusters of similar images and clean boundaries between them. To further increase the results, a pseudo-labelling approach is employed to take the most confident model predictions and use them to improve the model, modelling both unsupervised and semi-supervised learning methods simultaneously. CCA is evaluated against the HAM10000 dataset measured in cluster purity, silhouette score, and expert re-view. The results demonstrate that CCA can cluster skin lesions with high accuracy and consistency without the need for labeled data. The model's stability, the level of confidence in the results, and the expectations in the medical field indicate that CCA has tremendous potential for skin diagnosis-related applications where labelled data cannot be obtained.
