Purpose: Simulating lower-dose exams from existing exams is an effective approach to image quality assessment and dose optimization in clinical CT. Currently, most simulation techniques are in projection domain to incorporate appropriate noise models, which is inconvenient and time consuming. The purpose of this work was to develop a deep convolutional neural network (CNN) that synthesizes realistic low-dose patient CT exams directly in image domain.
Methods: The proposed CNN was constructed with the in-house-developed network architecture. The design of the network architecture followed a generalized mathematical model that enabled the synthesis of lower-dose CT images, using the routine-dose CT images and Gaussian white noise as the inputs. The objective function of the CNN jointly minimized three customized loss functions: a perceptual loss function to achieve perceptually-realistic low-dose CT images; a frequency-spectrum loss to quantitatively match the noise frequency components; a diversity loss to ensure sufficient diversity of noise realization. The CNN was trained and tested using the patient CT exams from AAPM low-dose CT grand challenge (training / testing cases: 10 / 20). The noise texture was qualitatively assessed by 3 board-certified medical physicists. The image noise was quantitatively compared in the liver parenchyma region of patient images. Noise power spectrum (NPS) was also measured using the repeated scans of different-sized water phantom.
Results: The CNN-synthesized low-dose patient images had perceptually-comparable noise texture to that of the reference images. In patientsâ€™ liver parenchyma, the mean absolute percent difference of the noise level was < 3.0%. The CNN yielded comparable NPS curves to those measured from real low-dose scans across different phantom sizes (the mean absolute difference < 1.1 HUÂ²2cmÂ²).
Conclusion: The proposed CNN method demonstrated the potential of simulating realistic lower-dose patient CT exams directly in image domain, which is more efficient and convenient than conventional projection-domain approaches.