Room: Exhibit Hall | Forum 8
Purpose: In the past several years, model-based image reconstruction (MBIR) methods were equipped with multi-detector CT (MDCT) scanners to enable low dose CT imaging. However, several challenges have been identified in image quality assessment including peculiar characteristics in spatial resolution, noise power spectrum, and CT number accuracy. These inferior characteristics are partially attributed to the use of spatial regularizers in current MBIR methods. In this work, a novel sandwich architecture including unregularized MBIR at two ends with a deep convolutional neural network (CNN) in between was developed to reconstruct low dose CT images.
Methods: The proposed framework (Fig. 1) starts with a convergent unregularized MBIR reconstruction from low dose projection data. The reconstructed noisy images were denoised by a deep CNN, followed by another unregularized MBIR reconstruction step. The design rationale for the first MBIR reconstruction was to eliminate structured noise streaks in low dose CT reconstruction; the CNN was then used to denoise; and the final MBIR reconstruction step was to correct potential spatial resolution loss and un-natural noise texture due to the CNN step. The proposed CNN consists of six 3Ã—3 convolutional layers with ReLu activations. A concatenation and residual learning strategy are leveraged to preserve spatial resolution and boost training performance. Low dose CT projection data were generated from standard dose CT exams acquired from MDCT scanner (Discovery CT 750HD, GE Healthcare) with Poisson noise added to simulate low dose CT scans.
Results: Noise standard deviations measured at eight ROIs from standard dose FBP images, low dose FBP images and images reconstructed by the proposed framework are 43Â±5 HU, 150Â±28 HU and 55Â±11 HU.
Conclusion: Final images reconstructed by the proposed sandwich architecture demonstrate greatly mitigated noise streaks, reduced uniform noise amplitude and relatively natural noise texture.