Less is more: Few-shot learning in 3D super resolution using synthetic CT images

Abstract

In this work, we study data requirements and scaling behavior of super resolution (SR) machine learning networks on computed tomography (CT) scans of aluminum foam samples. The pipeline consists of pre-training a 3D U-Net model on purely synthetic data and fine-tuning this model on increasing quantities of real data. Synthetic data is produced by a shallow convolutional neural network (CNN), which learns real CT degradations. This degraded data is later used for training a 3D U-Net SR network in order to reduce manual alignment labor for training such models. Our results suggest that only a small fraction of the dataset is required for fine-tuning synthetic models and diminishing returns are observed after a few pairs. Furthermore, simply increasing the amount of real fine-tuning data does not consistently improve performance across the different data configurations. These findings indicate that dataset selection may be as important as dataset size when training SR models for CT enhancement.