Patchwise-connected resolution enhancement for healthy weight-bearing bones using deep learning

Abstract

As our society is aging, osteoporosis is becoming prevalent. Accurate diagnosis of osteoporosis needs a high-resolution skeletal image. Resolution enhancement for skeletal images using topology optimization is successful but the problem is that its computational time is too long for clinical application. Recently, deep learning receives attention due to its high accuracy and fast computational time, therefore, resolution enhancement using deep learning has been researched. However, a lack of large-scale datasets is a major problem for deep learning in medical images. The patchwise approach can solve this problem to augment many data, however, this approach causes disconnected patches, which degrades image quality and structural behavior. This paper introduces quilting algorithm, which is used in texture synthesis, to solve this discontinuity problem. In addition, ResNet and SRGAN, which are commonly used in resolution enhancement problem, are compared in image quality and structural behavior. The proposed method reconstructs high-resolution skeletal images faster than topology optimization and improves image quality and structural behavior. However, the accuracy is still a lot lower, so further works are needed.