Comparison of self-supervised and ImageNet pretraining for medical image classification

Deep learning has witnessed a huge amount of attention over the last decade.It has been widely used in the field of medical images, especialy the deep neural networks for classification, detection and segmentation task. However, when compared with natural images, medical images are very difficult to label. In most cases, the size of the data is difficult to match the natural image database.To address this problem, the traditional approach is through the use of transfer learning methods, by using ImageNet pre-trained models. But medical images are often very different from natural images, the light and dark of the images have special meanings, and the images contain internal structures of the human body that are completely absent from the natural image data set. In this thesis, for 21 datasets, we used a self-supervised learning approach to train medical images of different modalities(CT, MRI, X-RAY) together into a neural network, applying the pre-trained weights to downstream medical image datasets for the classification task, we trained three different models, training from scratch, transfer learning by using self-supervised pre-trained weights, transfer learning by using ImageNet weights, and analyzing the differences between the three models by comparing the three models AUC (Area Under the Curve).