A Comparative Analysis of Transfer Learning Methods for Medical Super-Resolution

Abstract

Precisely and rapidly acquiring high-resolution medical images with a limited radiation dose is of great importance in clinical diagnosis. However, due to the challenges in gathering medical images and the high costs of developing a deep learning model from scratch, the super-resolution of medical images is still a tough problem. To deal with above, transfer learning is a favorable technique that leverages a previously trained model for a different but related task, significantly cutting down on the amount of training data and time needed for the downstream task. In this study, a pre-trained U-shaped Network++ (UNet++) model is initially trained using the ImageNet dataset to learning the general information. Subsequently, multiple finetuning methods are adopted to transfer the natural images super-resolution to the super-resolution of skin lesion images. 1279 manually annotated skin lesion images are selected for training and testing. The results indicate that transfer learning not only improve the convergence speed of deep learning models but also generate clearer high-resolution skin lesion images. Compared to the normal training group, the Mean Square Error (MSE) value was dropped by 24% to 70%. Notably, BitFit might lead to worse performance.