engleski

Optimization of SPECT procedures via Deep Convolutional Neural Network

Purpose. Nuclear medicine imaging procedures are characterized by long acquisition times. The number of acquired projections correlates with the acquisition time and reconstructed image quality. Therefore, acquisition time is parameter to be optimized during comprehensive optimization process. This study assesses the feasibility of using a deep convolutional neural network (DCNN) to improve the reconstructed SPECT image quality. The image quality similar to the one obtained with twice the number of planar projections should be achieved, which in turn implies halved acquisition time. Materials and Methods: The DCNN was implemented in the PyTorch and trained using 36 acquisitions of the multiple head registration/center of rotation (MHR/COR) phantom and 72 acquisitions of the Jaszczak phantom with, overall, 6768 projections. The network is to produce a synthesized projection that is to approximate the missing n-th projection given a pair of (n−1, n+1) acquired projections. Synthesized projections were compared to projections from a baseline method, which generated the missing projection with a pixel-by- pixel arithmetic mean of neighbour projections. The obtained synthesized projections were compared across several similarity measures: Structural Similarity Index Measure (SSIM), Visual Information Fidelity (VIF), Haar Perceptual Similarity Index (HaarPSI) and Mean Absolute Error (MAE) . The final augmented acquisition composed of acquired and synthesized projections was reconstructed using an iterative reconstruction algorithm. Its tomographic spatial resolution and contrast were compared to the reconstruction of the original image and the baseline image. The FWHM of 5 point-spread functions of MHR/COR phantom were compared for reconstructed original, DCNN generated, and baseline images of MHR/COR phantom generated as described for Jaszczak phantom. Results. DCNN synthesized projections achieved higher SIM, VIF, HaarPSI and MAE scores when compared with original and baseline projections. Augmented images of Jaszczak phantom achieved better tomographic image quality results than baseline images or acquired images. The image quality parameters of augmented image were comparable to the tomographic images reconstructed from real acquisitions with twice as many projections. Analysis and comparison of results for MHR/COR phantom showed that the DCNN could produce images with spatial resolution better than baseline and comparable to the original acquisitions. Conclusions. The study demonstrates the positive results in applying DCNN trained on Jaszczak and MHR/COR phantom images: the improvements in reconstructed SPECT image quality and results comparable to acquisitions with twice as many projections. The application of deep learning in nuclear medicine shows a clear potential in acquisition time optimization process with retaining satisfactory image quality results.

SPECT, optimization, image quality, deep convolutional neural network

nije evidentirano