Disease Diagnosis Based on Deep Learning and Transfer Learning: Evolution from CNN to SAM

Medical image analysis is crucial for clinical diagnosis. Recent advances in deep learning, especially convolutional neural networks (CNNs), have driven progress in this field. However, challenges such as high annotation costs, small sample sizes, and cross-center data heterogeneity limit model performance. Transfer learning and pre-trained models have therefore become key solutions to these issues. This article with a review of medical image deep learning as the thread, systematically sorts out the core types and mechanisms of transfer learning, the evolution and lightweight path of classic CNN pre-training models, explores the technological development trends from CNN to Vision Transformers (ViT) and Segment Anything Model (SAM), and conducts a comprehensive analysis and discussion. Finally, it summarizes the key challenges and future research directions in the field, clarifies the technological evolution rules in this field, and provides important theoretical and practical references for building clinically applicable medical imaging intelligent systems.