MedUnA: Efficient Medical Image Classification through Unsupervised Adaptation of Vision-Language Models

MedUnA, developed by researchers from Mohamed Bin Zayed University of AI and Inception Institute of AI, improves medical image classification using unsupervised adaptation of Vision-Language Models (VLMs). The method employs a two-stage training process: adapter pre-training with text descriptions and unsupervised learning using medical images. This approach reduces the reliance on labeled data and enhances scalability, offering improved efficiency and performance without extensive pre-training. Experiments on multiple medical datasets showed that MedUnA outperforms existing methods, such as CLIP and MedCLIP, in classification accuracy.