Investigation of Direct and Indirect Approaches of Deep-Learning-Based Attenuation Correction for General Purpose and Dedicated Cardiac SPECT Scanners

Attenuation correction using CT transmission scanning enables accurate quantitative assessment of cardiac SPECT. Deep-learning-based indirect approaches have been established to predict attenuation maps from emission data for rotational SPECT-only scanners with parallel-hole collimators and NaI crystals. Direct transformation approaches to generate attenuation-corrected images from non-attenuation-corrected images might be easier to implement without the intermediate step of the attenuation map generation, particularly useful for the small field-of-view of dedicated cardiac SPECT scanners with CZT detectors. In this work, we first implemented and compared the direct and indirect approaches for both conventional parallel-hole SPECT using 200 anonymized datasets and dedicated cardiac pinhole SPECT using 176 anonymized datasets. To avoid the inaccuracy caused by truncated reconstruction of the dedicated SPECT, we proposed novel methods to predict truncated attenuation maps from truncated emission images, and full attenuation maps from full though inaccurate emission images. The predicted truncated and full attenuation maps were then zero-padded and incorporated into the iterative reconstruction to generate attenuation-corrected images. For parallel-hole SPECT. the averaged error of the attenuation-corrected images using the direct approach was 2.57 ± 1.06% as compared to 1.37 ± 1.16% using the indirect approach. For the dedicated pinhole cardiac SPECT, the averaged error using our proposed indirect approaches was 1.14 ± 0.74% as compared to 2.20 ± 1.11% using the direct approach. In addition, we designed and implemented a novel neural network that can better extract information from a multi-channel input, which showed superior performance than conventional U-Net in both indirect and direct approaches.