Shorter SPECT scans using self-supervised coordinate learning to synthesize skipped projection views

Abstract

Abstract Purpose This study addresses the challenge of extended SPECT imaging duration under low-count conditions, as encountered in Lu-177 SPECT imaging, by developing a self-supervised learning approach to synthesize skipped SPECT projection views, thus shortening scan times in clinical settings. Methods We developed SpeRF, a SPECT reconstruction pipeline that integrates synthesized and measured projections, using a self-supervised coordinate-based learning framework inspired by Neural Radiance Fields (NeRF). For each single scan, SpeRF independently trains a multi-layer perceptron (MLP) to estimate skipped SPECT projection views. SpeRF was tested with various down-sampling factors (DFs = 2, 4, 8) on both Lu-177 phantom SPECT/CT measurements and clinical SPECT/CT datasets, from 11 patients undergoing [177Lu]Lu-DOTATATE and 6 patients undergoing [177Lu]Lu-PSMA-617 radiopharmaceutical therapy. Performance was evaluated both in projection space and by comparing reconstructed images using (1) all measured views (“Full”), (2) down-sampled measured views only (“Partial”), and partially measured views combined with skipped views (3) generated by linear interpolation (“LinInt”) and (4) synthesized by our method (“SpeRF”). Results SpeRF projections demonstrated lower Normalized Root Mean Squared Difference (NRMSD) compared to the measured projections, than LinInt projections. Across various DFs, the NRMSD for SpeRF projections averaged 7% vs. 10% in phantom studies, 18% vs. 26% in DOTATATE patient studies, and 20% vs. 21% in PSMA-617 patient studies, compared to LinInt projections. For SPECT reconstructions, DF = 4 is recommended as the best trade-off between acquisition time and image quality. At DF = 4, in terms of Contrast-to-Noise Ratio relative to Full, SpeRF outperformed LinInt and Partial: (1) DOTATATE: 88% vs. 69% vs. 69% for lesions and 88% vs. 73% vs. 67% for kidney, (2) PSMA-617: 78% vs. 71% vs. 69% for lesions and 78% vs. 57% vs. 67% for organs, including kidneys, lacrimal glands, parotid glands, and submandibular glands. SpeRF slightly underestimated count recovery relative to Full, compared to Partial but still outperformed LinInt: (1) DOTATATE: 98% vs. 100% vs. 88% for lesions and 98% vs. 100% vs. 94% for kidney, (2) PSMA-617: 98% vs. 101% vs. 94% for lesions and 96% vs. 101% vs. 78% for organs. Conclusion The proposed method, SpeRF, shows potential for significant reduction in acquisition time (up to a factor of 4) while maintaining quantitative accuracy in clinical SPECT protocols by allowing for the collection of fewer projections. The self-supervised nature of SpeRF, with data processed independently on each patient’s projection data, eliminates the need for extensive training datasets. The reduction in acquisition time is particularly relevant for imaging under low-count conditions and for protocols that require multiple-bed positions such as whole-body imaging.

Keywords

• Self-supervised learning
• SPECT
• Sparse projection views
• Lu-177 imaging