A geometry-guided multi-beamlet deep learning technique for CT reconstruction

Abstract

Purpose. Previous studies have proposed deep-learning techniques to reconstruct CT images from sinograms. However, these techniques employ large fully-connected (FC) layers for projection-to-image domain transformation, producing large models requiring substantial computation power, potentially exceeding the computation memory limit. Our previous work proposed a geometry-guided-deep-learning (GDL) technique for CBCT reconstruction that reduces model size and GPU memory consumption. This study further develops the technique and proposes a novel multi-beamlet deep learning (GMDL) technique of improved performance. The study compares the proposed technique with the FC layer-based deep learning (FCDL) method and the GDL technique through low-dose real-patient CT image reconstruction.Methods. Instead of using a large FC layer, the GMDL technique learns the projection-to-image domain transformation by constructing many small FC layers. 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 FC layers in GMDL connect each pixel to beamlets peripheral to the central beamlet based on the CT projection geometry. We compare ground truth images with low-dose images reconstructed with the GMDL, the FCDL, the GDL, and the conventional FBP methods. The images are quantitatively analyzed in terms of peak-signal-to-noise-ratio (PSNR), structural-similarity-index-measure (SSIM), and root-mean-square-error (RMSE).Results. Compared to other methods, the GMDL reconstructed low-dose CT images show improved image quality in terms of PSNR, SSIM, and RMSE. The optimal number of peripheral beamlets for the GMDL technique is two beamlets on each side of the central beamlet. The model size and memory consumption of the GMDL model is less than 1/100 of the FCDL model.Conclusion. Compared to the FCDL method, the GMDL technique is demonstrated to be able to reconstruct real patient low-dose CT images of improved image quality with significantly reduced model size and GPU memory requirement.

Keywords: CT; deep learning; fully connected layer; reconstruction.

Figure 1.

General workflow of the study

Figure 2.

Fan-beam CT projection geometry

Figure 3.

The network architecture of the (a) GMDL reconstruction module and (b) net block for a single projection

Figure 4.

Box plots of (a) PSNR, (b) SSIM, and (c) RMSE values calculated with both the large ROI and the small ROIs on images reconstructed with the FBP method and GDL, GMDL, and FCDL models over the two test sets

Figure 4.

Box plots of (a) PSNR, (b) SSIM, and (c) RMSE values calculated with both the large ROI and the small ROIs on images reconstructed with the FBP method and GDL, GMDL, and FCDL models over the two test sets

Figure 5.

CT images reconstructed by the FBP method and the GDL, GMDL, and FCDL models from two test sets and corresponding difference images. The squares marked by red and yellow lines in the ground truth are the large ROI and small ROIs from which the PSNR, SSIM, and RMSE values are calculated

Figure 5.

CT images reconstructed by the FBP method and the GDL, GMDL, and FCDL models from two test sets and corresponding difference images. The squares marked by red and yellow lines in the ground truth are the large ROI and small ROIs from which the PSNR, SSIM, and RMSE values are calculated

Figure 6.

CT images reconstructed by GMDL and FCDL models using the ACR phantom. The reconstructed CT image has a pixel size of 0.7422 mm × 0.7422 mm

Figure 7.

CT images reconstructed with FBP, 10 iterations of SART, and GMDL-2P

Figure 8.

Z-locations of outlier CT images reconstructed by FBP method, GDL, GMDL, and FCDL models in terms of (a), (b) PSNR, (c), (d) SSIM, and (e), (f) RMSE calculated on both the large ROI ((a), (c), and (e)) and small ROIs ((b), (d), and (f)) in the two test cases. The red area means upper outliers with PSNR, SSIM, or RMSE values greater than Q3 + 1.5IQR, where Q3 is the 75^th percentile, and IQR is the interquartile range. The green area means lower outliers with the corresponding values lower than Q1 − 1.5IQR, where Q1 is the 25^th percentile.

Figure 8.

Z-locations of outlier CT images reconstructed by FBP method, GDL, GMDL, and FCDL models in terms of (a), (b) PSNR, (c), (d) SSIM, and (e), (f) RMSE calculated on both the large ROI ((a), (c), and (e)) and small ROIs ((b), (d), and (f)) in the two test cases. The red area means upper outliers with PSNR, SSIM, or RMSE values greater than Q3 + 1.5IQR, where Q3 is the 75^th percentile, and IQR is the interquartile range. The green area means lower outliers with the corresponding values lower than Q1 − 1.5IQR, where Q1 is the 25^th percentile.

Figure 9.

Backprojection from a single projection using (a) GDL (GMDL-0P), (b) GMDL-2P and (c) FCDL models. The single projection has two bright pixels (one in the middle and the other on the side). The bright pixels have an intensity of 1, and all other dark pixels are set to 0. (d) The line profiles drawn along the gray dashed lines in (a), (b) and (c)

Figure 10.

CT images reconstructed with GMDL-2P with different pixel sizes. The left one has a pixel size of 0.7422 mm × 0.7422 mm while the one on the right has a pixel size of 0.664 mm × 0.664 mm