Artificial intelligence based SPECT attenuation correction. Feasibility study and performance evaluation on clinical SPECT images.

Introduction: Precise attenuation correction (AC) is crucial to increase the image quality and enables accurate quantification in single-photon emission computed tomography (SPECT) imaging. Non-uniform attenuation correction using computed tomography (CT) is the gold standard method in clinical routine. However, significant artifacts can appear and affect the accuracy of applied attenuation maps due to patient motion, misregistration between the SPECT and CT images, or the presence of metal implants and contrast agents in the human body. Furthermore, CT scans increase the radiation dose of patients as well. SPECT-only systems are still commonly used worldwide, which do not allow transmission imaging, therefore the option of CT-based AC is not available. We aimed to develop an artificial intelligence (AI) supported method which is able to generate synthetic CT images from bone SPECT measurements. This novel technique paves the way to achieve quantitative SPECT imaging similar to the results of the standard CT-based correction method without the need of an actual CT scan.

Methods: A convolutional neural network (CNN) based solution with supervised and unsupervised techniques was built by our group. The task of the neural network was to generate a synthetic CT-like image from the SPECT data and a corresponding attenuation map. During training, we aimed to make this result as similar as possible to the attenuation map obtained using the original CT volume acquired at the same time as the SPECT. The retrospective study included 1300 clinical subjects with both normal and pathological patients that injected with ^99mTc-MDP and images were acquired on AnyScan TRIO^® SPECT/CT system. Clinical sample was heterogeneous in terms of sex, age, BMI, disease and examined regions. Clinical images were separated by random selection: 900 were used for network training, 200 for testing and 200 for evaluation. For image processing and performance evaluation, the available SPECT projection data was reconstructed in three different ways: attenuation correction using the original CT volume (CT-AC), attenuation correction using the synthetic CT volume (SCT-AC) and non-attenuation-corrected (Non-AC). Both the attenuation maps and the reconstructed SPECT images were compared by considering the original CT-based method as the reference. Image-derived metrics including mean absolute error, structural similarity index (SSI), peak signal-to-noise ratio (SNR), Dice coefficient, and region-based clinically relevant values such as relative and absolute activity concentrations were evaluated.

Results: Evaluation of 200 independent samples (i.e., not involved in neural network learning and hyperparameter optimization) showed that the synthetic attenuation maps are qualitatively and quantitatively consistent with the CT-based attenuation maps. There was no significant difference in SNR and activity concentration between CT-AC and SCT-AC reconstructed SPECT volumes. As expected, both CT-AC and SCT-AC methods significantly outperformed Non-AC in terms of SNR. Furthermore, misregistration errors between SPECT and CT images caused a decrease in the accuracy of CT-AC SPECT images, whereas when using SCT-AC, misregistration is excluded due to the main principle of our method, and this type of error should never affect the resulting SPECT images.

Conclusions: In this study, we have demonstrated with our initial experimental results that accurate attenuation maps can be generated only from the emission data for bone SPECT imaging. The proposed method has several clinical advantages such as reducing the radiation dose to the patient and eliminating CT and motion artifacts from SPECT images.