Soft-tissue radiography using multi-scale convolutional neural networks

Abstract

Dual-energy imaging techniques based on the fast kVp switching and the sandwich detector can enhance the conspicuity in digital radiography. Deep learning is an alternative promising approach to obtain dual-energy images without hardware cost and additional patient dose. This paper extends the well-known multi-scale U-net to preserve high-frequency information so as to be applicable to structural filtering of medical images with minimal loss of information. The performance of modified versions of the U-net is presented in comparison with the conventional U-net architecture. Limitation of the modified networks is also discussed.