Deep Learning-based CBCT Projection Interpolation, Reconstruction, and Post-processing for Radiation Therapy
Access is limited until:
Cone-beam computed tomography (CBCT) is an X-ray-based imaging modality widely used in medical practices. Due to the ionizing imaging dose induced by CBCT, many studies were conducted to reduce the number of projections (sparse sampling) to lower the imaging dose while maintaining good image quality and fast reconstruction speed. Conventionally, a CBCT volume is reconstructed analytically with the Feldkamp Davis Kress (FDK) algorithm that backprojects filtered projections according to projection angles. However, the FDK algorithm requires a dense angular sampling that satisfies the Shannon-Nyquist theorem. The FDK algorithm reconstructs CBCT with a high speed but requires relatively high patient imaging dose. The iterative methods like algebraic reconstruction technique (ART) and compressed sensing (CS) methods are investigated to reduce patient imaging dose. These iterative methods update estimated images iteratively and the CS methods apply penalty terms to award desired features. Yet these methods are limited by the iterative design with substantially increased computation time and consumption of computation power. Scholars have also conducted research on bypassing the limit of Shannon-Nyquist theorem by interpolating densely sampled CBCT projections from sparsely sampled projections. However, blurred structures in reconstructed images remain to be a concern for analytical interpolation methods. As such, previous research indicates that it is hard to achieve the three goals of lowered patient imaging dose, good image quality, and fast reconstruction speed all at once.
As deep learning (DL) gained popularity in fields like computer vision and data science, scholars also applied DL techniques in medical image processing. Studies on DL-based CT image reconstruction have yielded encouraging results, but GPU memory limitation made it challenging to apply DL techniques on CBCT reconstruction.
In this dissertation, we hypothesize that the image quality of CBCT reconstructed from under-sampled projections (low-dose) using deep learning techniques can be comparable to that of CBCT reconstructed from fully sampled projections for treatment verification in radiation therapy. This dissertation proposes that by applying DL techniques in pre-processing, reconstruction, and post-processing stages, the challenge of improving CBCT image quality with low imaging dose and fast reconstruction speed can be mitigated.
The dissertation proposed a geometry-guided deep learning (GDL) technique, which is as the first technique to perform end-to-end CBCT reconstruction from sparsely sampled projections and demonstrated its feasibility for CBCT reconstruction from real patient projection data. In this study, we have found that incorporating geometry information into the DL technique can effectively reduce the model size, mitigating the memory limitation in CBCT reconstruction. The novel GDL technique is composed of a GDL reconstruction module and a post-processing module. The GDL reconstruction module learns and performs projection-to-image domain transformation by replacing the traditional single fully connected layer with an array of small fully connected layers in the network architecture based on the projection geometry. The additional deep learning post-processing module further improves image quality after reconstruction.
This dissertation further optimizes the number of beamlets used in the GDL technique through a geometry-guided multi-beamlet deep learning (GMDL) technique. In addition to connecting each pixel in the projection domain to beamlet points along the central beamlet in the image domain as GDL does, these smaller fully connected layers in GMDL connect each pixel to beamlets peripheral to the central beamlet based on the CT projection geometry. Due to the limitation of GPU VRAM, the proposed technique is demonstrated through low-dose CT image reconstruction and is compared with the GDL technique and a large fully connected layer-based reconstruction method.
In addition, the dissertation also investigates deep learning-based CBCT projection interpolation and proposes a patient-independent deep learning projection interpolation technique for CBCT reconstruction. Different from previous studies that interpolate phantom or simulated data, the proposed technique is demonstrated to work on real patient projection data with unevenly distributed projection angles. The proposed technique re-slices the stack of interpolated projections axially, and each acquired slice is processed by a deep residual U-Net (DRU) model to augment the slice’s image quality. The resulting slices are reassembled into a stack of densely-sampled projections to be reconstructed into a CBCT volume. A second DRU model further post-processes the reconstructed CBCT volume to improve the image quality.
In summary, a geometry-guided deep learning (GDL) technique was proposed as the first deep learning technique for end-to-end CBCT reconstruction from sparsely sampled real patient projection data. The geometry-guided multi-beamlet deep learning (GMDL) technique further optimizes the number of beamlets based on the GDL technique. A patient-independent deep learning projection interpolation technique was also proposed for the pre-processing and post-processing stage of CBCT reconstruction.
In conclusion, the work presented in this dissertation demonstrates the feasibility of improving CBCT image quality with low imaging dose and fast reconstruction speed. The techniques developed in this dissertation also have great potential for clinical applications to enhance CBCT imaging for radiation therapy.
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License.
Rights for Collection: Duke Dissertations
Works are deposited here by their authors, and represent their research and opinions, not that of Duke University. Some materials and descriptions may include offensive content. More info