The Evolution of Deep Learning in Medical Imaging

Abstract

In the past 10 years, the evolution path of medical imaging AI in terms of model architecture is very clear: it is evolving from CNNs for end-to-end feature learning, to ResNets for alleviating deep degradation issues, to EfficientNet balancing accuracy and efficiency with compound scaling, to Transformers for global modeling with self-attention. It is systematic and clear, but it is also important to map it systematically, which will help to be clearer about technology selection and further optimize clinical translation. This paper focuses on four major models/captures CNN, ResNet, EfficientNet, and Transformer, summarizes their basic designs and innovation points.