Cross-Tracer and Cross-Scanner Transfer Learning-Based Attenuation Correction for Brain SPECT

Abstract

This study aims to investigate robust attenuation correction (AC) by generating attenuation maps $(\mu $ -maps) from nonattenuation-corrected (NAC) brain SPECT data using transfer learning (TL). Four sets of brain SPECT data ( $4\times 30$ ) were retrospectively collected: S-TRODAT-1, S-ECD, G-TRODAT-1, and G-ECD. A 3-D attention-based conditional generative adversarial network was pretrained using 22 paired 3-D NAC SPECT images and corresponding CT $\mu $ -maps for four patient groups. Various numbers ( $n\,\,=$ 4–22) of paired NAC SPECT and corresponding $\mu $ -maps from S-TRODAT-1 were then used to fine-tune (FT) the other three pretrained deep learning (DL) networks, i.e., S-ECD, G-TRODAT-1, and G-ECD. All patients in S-TRODAT-1 group were tested on their own network (DL-AC), and on the pretrained models with FT (FT-AC) and without FT (NFT-AC). The FT-AC methods used 22 (FT22), 12 (FT12), 8 (FT8), and 4 (FT4) paired data for FT, respectively. Our results show that FT22 and FT12 could outperform DL-AC for cross-tracer S-ECD and cross-scanner G-TRODAT-1 using CT-based AC (CT-AC) as the reference. FT22 also outperforms DL-AC for cross-tracer+cross-scanner G-ECD. FT8 performs comparably to DL-AC, while FT4 is worse than DL-AC but still better than NAC and NFT-AC in each group. Attenuation map generation is feasible for brain SPECT based on cross-tracer and/or cross-scanner FT-AC using a smaller number of patient data. The FT-AC performance improves as the number of data used for FT increases.