DeepSPECT: A Deep Learning Based Framework for Fast SPECT Imaging

Purpose: The purpose of this work was to assess the feasibility of reducing the acquisition time for single-photon emission computed tomography (SPECT) imaging using deep learning based framework without comprising the image diagnostic quality. Method and Materials: 20 subjects received systemic bone imaging and quantitative tomography were collected on a SPECT / CT scanner (Siemens Symbia Intevo) at Shanghai East Hospital (Dosage of injection, 25 mCi). We set two scanning protocols: one standard scan with 20 seconds per frame (referred as standard SPECT), and one fast scan with 3 seconds per frame (referred as 1/7 SPECT). Other sampling parameters were as follows: 60 frames, single probe rotated 180 °, single rotation 6 °. SPECT projection data were reconstructed using ordered subset conjugate gradient (OSCG) algorithm. The image-to-image transformation via the U²-Net was implemented to predict standard SPECT image from 1/7 SPECT image and corresponding CT image, as shown in Fig. 1. The main architecture is a U-Net like Encoder-Decoder, where each stage consists of our newly proposed residual U-block (RSU). Due to the mixture of receptive fields of different sizes in our proposed RSU, the model is able to capture local and global information from both shallow and deep layers regardless of the resolutions. This novel architecture also increases the depth of the whole network without significantly increasing the computational cost.10 subjects were chosen for training this deep learning model. The remaining 10 subjects are used for testing model performance. Qualitative and quantitative evaluations of the proposed framework was performed using structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) metrics.

Results: The results demonstrated that the DL recovered SPECT had image quality comparable to standard SPECT images, much better than the original 1/7 SPECT images as shown in Fig. 2. It is relatively difficult to identify the lesion regions on 1/7 SPECT images, while much easier on the recovered SPECT image using proposed deep learning model. Quantitative assessments were shown in Table 1. Conclusion: Reducing the acquisition time for SPECT scan significantly degrade the image quality (low PSNR and low SSIM). The proposed deep learning based method can effectively recover and improve image quality with quantification measurements comparable to standard SPECT scan. Further research is granted to explore the diagnostic accuracy using 1/7 SPECT image, recovered SPECT image and standard SPECT image respectively.

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint