Semi-stationary Multi-source AI-powered Real-time Tomography

Abstract

Over the past decades, the development of CT technologies has been largely driven by the need for cardiac imaging but the temporal resolution remains insufficient for clinical CT in difficult cases and rather challenging for preclinical CT since small animals have much higher heart rates than humans. To address this challenge, here we report a Semi-stationary Multi-source AI-based Real-time Tomography (SMART) CT system. This unique scanner is featured by 29 source-detector pairs fixed on a circular track to collect X-ray signals in parallel, enabling instantaneous tomography in principle. Given the multisource architecture, the field-of-view only covers a cardiac region. To solve the interior problem, an AI-empowered interior tomography approach is developed to synergize sparsity-based regularization and learning-based reconstruction. To demonstrate the performance and utilities of the SMART system, extensive results are obtained in physical phantom experiments and animal studies, including dead and live rats as well as live rabbits. The reconstructed volumetric images convincingly demonstrate the merits of the SMART system using the AI-empowered interior tomography approach, enabling cardiac CT with the unprecedented temporal resolution of 33ms, which enjoys the highest temporal resolution than the state of the art.

Keywords: Computed tomography (CT); cardiac imaging; deep learning; image reconstruction; multi-source; preclinical imaging; real-time.

Figure 1.

Multi-source “SMART” preclinical CT prototype and reconstruction framework for dyamic image reconstruction. (a) A photograph of the real system; (b) the imaging geometry; (c) the flowchart of the SMART reconstruction framework; (d) the interior CT network in the first SMART reconstruction stage; (e) animal model for contrast injection with dynamic imaging; (f) training, validation and testing for deep neural network-based interior reconstruction.

Figure 2.

Volumetric rendering of an extracted lung region of interest using different reconstruction algorithms. (a) and (b) The reference from a full projection dataset without truncation using the TDL algorithm, (c) and (d), (e) and (f), (g) and (h) are the images reconstructed from only 29 views using FDK, TVM and our AI-empowered interior tomography methods respectively. (b), (d), (f) and (h) represent the extracted lung region to further demonstrate AI-empowered interior tomography approach advantages.

Figure 3.

Representative transverse slices through a dead rat to validate our AI-empowered interior tomography reconstruction network. (a) The reference from a full projection dataset without truncation using the TDL algorithm. (b)-(d) The images reconstructed from only 29 views using FDK, TVM and our methods respectively. The images of (b), (c) and (d) from sparse data using different methods demonstrate remarkable image quality variations. The images of (b) column is unacceptable due to strong artifacts, poor texture and inability to assess some anatomical structures. The images of (c) column have much-reduced artifacts but look smooth, compromising texture for clinical usability. The images of (d) column using our approach have optimal image quality in terms of texture conservation (yellow circle), artifact suppression, and clear visualization of small structures (red circle). The display window is [0 0.065] in terms of the linear attenuation coefficient.

Figure 4.

Reconstruction results of the alive rat to evaluate our reconstruction technique. (a)-(c) contain three images from transverse, coronal and sagittal reconstructed from only 29 views using FDK, TVM, and our methods, respectively. The display window is [0 0.065] in terms of the linear attenuation coefficient. The images of (a) with FDK reconstruction are unacceptable for strong artifacts, poor texture and undermined structures. The images from (b) are significantly better but they are smooth and blocky. The images of (c) reconstructed with our method have optimal image quality. The display window is [0 0.065] in terms of the linear attenuation coefficient.

Figure 5.

Sequential CT images of the alive rat in three timeframes to visualize dynamic changes with extracted tissue. The images in (a-c), (d-f) and (g-i) are reconstructed using the FDK, TVM, and AI-empowered interior tomography methods, respectively. The display window is [0 0.065] in terms of the linear attenuation coefficient.

Figure 6.

CT images of the alive rabbit at one time frame reconstructed from 29 projections. The transverse, coronal and sagittal images in (a)-(b), (c)-(d) and (e)-(f) were reconstructed using the FDK, TVM, and our reconstruction algorithms, respectively. The images (a), (c) and (e) are volume rendering images from FDK, TVM and our proposed methods. (b), (d) and (f) are transverse, sagittal and coronal images from FDK, TVM and AI-empowered interior tomography methods. The display window of (b), (d) and (f) is [0 0.065] in terms of the linear attenuation coefficient.

Figure 7.

Sequential CT images of the alive rabbit at four time instants to visualize the real-time cardiac dynamics from only 29 projections. The images in (a), (b) and (c) were reconstructed using the FDK, TVM and AI-empowered interior tomography methods, respectively. The images in (d) were profiles indicated by the blue position in (c) with our reconstruction method. The changes of thin blood vessels can be found in (c) but they cannot be seen in FDK and TVM results. It clearly shows that the system resolution reaches 33ms resolution, which is the highest imaging speed of CT ever reported.

Figure 8.

Dynamic CT perfusion. With 33ms temporal resolution enabled by our AI-empowered interior tomography approach, the perfusion process can be observed in real-time. The 1^st row shows projections in two different time instants, and the 2^nd row presents the corresponding reconstruction results respectively.