Room: Exhibit Hall | Forum 9
Purpose: Recently, studies on the development of low-dose computed tomography (CT) have been performed to reduce radiation. In particular, low-dose CT reconstruction based on deep learning attracts a significant attention as a next-generation technology to overcome the limitation of the conventional iterative reconstruction. We aim to verify whether the deep-learning-based low-dose CT image processing method can be generally used for clinical practice, and quantitatively analyze the image quality.
Methods: We used public medical imaging database for developing deep learning based image processing. 15 patients CT data were used for training and 3 patients CT data were used for testing. For noise CT images, we generated CT noise with computer simulation and added it to original image data. We used deep learning architecture based on U-net. U-net has characteristics that compensate for feature losses when dimension increases occur through the convolution transpose process, however still U-net has low frequency emphasized problem which resulted in image blurred in predicted image. In this study, we proposed novel architecture that improves resolution losses problem in U-net. To evaluate result images quantitatively, we used modulation transfer function (MTF) and noise power spectrum (NPS).
Results: Both U-net and proposed network showed excellent noise removing effect. The MTF value of proposed network is higher than U-net. There is slight loss of MTF in predicted images by both U-net and proposed network. MTF 50% of proposed network is approximately 10% higher than U-net, quantitatively. The NPS of proposed network and U-net is lower than high dose image. This means that, image characteristics of predicted image was superior to standard dose CT image in aspect of image noise.
Conclusion: Although a further training process with various clinical data is required to develop more sophisticated and practically usable algorithms, this study demonstrated the clinical applicability of deep-learning-based medical image processing.