Self-supervised Feature Representation Distillation for Esophageal Cancer Screening

Esophageal cancer is ranked among the most severe types of cancer and is a significant cause of cancer-related deaths. Recently, artificial intelligence (AI) techniques have been extensively employed for esophageal disease screening. However, the performance of these approaches is mainly guaranteed by the availability of large-scale labeled data, while collecting such data remains a significant challenge under clinical scenarios. In this paper, we propose a novel Self-supervised Feature Representation Distillation (SFRD) approach for the screening. The method begins by training a teacher model on labeled endoscopic esophageal images, followed by self-supervised learning to train a student model on unlabeled esophageal images. Knowledge distillation is then used to enhance the student model’s ability to extract discriminative features. Lastly, the student model is fine-tuned on labeled endoscopic esophageal images to further refine its feature extraction capabilities. Our experimental results on a clinical dataset demonstrate that the proposed SFRD method outperforms other competitive supervised and self-supervised approaches for esophageal cancer screening.