CLASH: A Contrastive Learning Approach for Few-Shot Classification of Histopathological Images

Real-world datasets often grapple with the challenge of long-tail phenomena and the scarcity of high-quality annotated images. Deep Learning (DL) models rely on vast datasets to enhance generalization. However, obtaining large datasets in domains such as the medical industry, defense, and security is challenging. Several studies have addressed this limitation by improving the DL model’s performance on small datasets. In this context, this study learns feature embeddings by employing a Siamese Network-based Contrastive Learning Approach for Few-Shot Classification of Histopathological images (CLASH), which consists of two sister networks of Convolutional Neural Networks (CNN). The model employs contrastive loss learning, ensuring that similar image pairs are closely mapped while dissimilar image pairs are distinctly separated. The small medical dataset of a few samples is used to fine-tune the contrastive learning model and learn the data representations for classification. The proposed model demonstrates exceptional performance even with as few as five samples per class. By leveraging the principles of few-shot learning (FSL), we affirm the model’s ability to extract relevant knowledge from a minimal dataset of few samples, showcasing its adaptability and performance in constrained training samples. Specifically, the model exhibits an impressive accuracy of 94% when trained on a dataset comprising merely fifty samples (FSL-50) per class. This underscores the model’s efficacy in addressing the challenges associated with small datasets and highlights its potential for practical applications in the medical domain.