Uni-Mlip: Unified Self-supervision for Medical Vision Language Pre-training

Vision-and-Language Pre-training (VLP) techniques, such as the Contrastive Language Image Pre-training (CLIP) model [35], have set a foundational approach for leveraging selfsupervision with language guidance to integrate visual and textual data effectively. By aligning image and text representations through contrastive learning, CLIP improves the model's ability to interpret complex data by linking visual concepts to linguistic descriptions. This method significantly enhances the understanding of multimodal representations by pre-training on extensive datasets and subsequently fine-tuning on targeted downstream tasks. This paradigm shift is particularly vital in the medical domain, where the acquisition of multimodal medical data presents significant challenges due to concerns over data privacy, sensitivity, and the complex, domain-specific knowledge required for annotation. Medical Vision-and-Language Pre-training (Med-VLP) aims to address such challenges inherent to the medical imaging field.