Reconstruction-Informed and Multidomain Deep Learning for Generalizable CT-Free Attenuation Correction in SPECT Myocardial Perfusion Imaging

Paper

Link to Source: Paper, Github

Authors: Ghasem Hajianfar, Yazdan Salimi, Mehdi Amini, Xiaotong Hong, René Nkoulou, Elnaz Jenabi, Zahra Mansouri, Atena Aghaee, Soroush Bagheri, Amirhossein Sanaat, Ahmad Bitarafan-Rajabi, Hossein Arabi, Isaac Shiri, Habib Zaidi

Summary: This study presents a deep learning approach that feeds several differently reconstructed non-corrected SPECT images into a single model and trains it against both the attenuation map and the corrected image at once, producing attenuation-corrected myocardial perfusion scans that hold up on data from an outside centre and that experienced readers could not reliably tell apart from the CT-corrected versions they were meant to replace.

Representative results illustrating the performance of deep learning models in generating ATM and AC images across various reconstruction inputs and training loss functions. Columns indicate input/output configurations, while rows reflect loss types: ATM loss, AC loss, combined ATM+AC loss, and Direct loss. Each panel includes the generated image (Gen Img) and the corresponding bias map. The bottom-left panel displays the reference images (Ref Img), which includes ATM, AC, and NAC images. ATM: Attenuation Map, AC: Attenuation-Corrected, NAC: non-AC, OSEM: Ordered Subset Expectation Maximization, 4I4S: OSEM with 4 Iterations and 4 Subsets. OSEM 3: combine 3 OSEMs.

Purpose

Deep learning (DL) has shown promise in enabling attenuation correction (AC) for SPECT myocardial perfusion imaging (MPI) without relying on anatomical information or CT-derived attenuation maps (ATMs). We introduce a novel reconstruction-informed and multidomain (RIMD) DL framework utilizing both multi-input reconstruction and dual-domain supervision for AC in SPECT MPI. Our method incorporates multi-input non-AC (NAC) images as input to the DL model and employs a combined loss function that optimizes performance in both the ATM and AC domains.

Methods

A dataset of 1058 SPECT/CT MPI scans using 99mTc-Sestamibi from two centers was used for training (934 cases) and an external test set (124 cases). SPECT projections were reconstructed into AC and NAC images using three reconstruction settings of the Ordered Subset – Expectation Maximization (OSEM) algorithm. SwinUnetR model was trained in a consistent 5-fold cross-validation framework with normalized NAC images as input. Our proposed method, as an indirect strategy, uses multi-input NAC images (incorporating all three images with different reconstruction settings) trained using a combination of ATM loss (between predicted and true ATMs) and AC loss (between AC images reconstructed from predicted ATMs and reference AC images). Evaluation included voxel-wise and region-wise metrics for both ATM and AC domains, 17-segment polar map analysis, and an organ-specific analysis. Clinical validation was also performed on part of the external dataset.

Results

Our proposed approach significantly outperformed direct and indirect methods in ablation comparison. This model yielded mean relative absolute error percentage (MRAE%) values of 25.02 ± 23 (internal) and 26.31 ± 14 (external) for ATMs, and 11.72 ± 6.3 (internal) and 19.31 ± 4.9 (external) for AC SPECT images. Organ-wise analysis showed region-wise MRAE% of 9.29 ± 6.5 (internal) and 17.28 ± 17 (external) in the ATM domain, and 4.51 ± 4.3 (internal) and 9.53 ± 6.6 (external) in the AC domain. Polar map analysis across 17 segments showed MRAE% of 5.04 ± 4.6 (internal) and 10.88 ± 7.3 (external). Clinical validation demonstrated high agreement between DLAC and CTAC images (ICC = 0.98), with physicians unable to distinguish between them (F1 score = 0.40), and no significant difference in diagnostic accuracy (DLAC: 0.63, CTAC: 0.70; p = 0.37).

Conclusion

This study demonstrated that our proposed RIMD method utilizing multiple OSEM reconstruction inputs and jointly optimizing ATM and AC losses substantially improved model performance in the indirect strategy. The indirect method consistently outperformed the direct approach, and our model generalized well on external data, showing strong agreement with SPECT CTAC images in both quantitative and qualitative assessments. Preliminary clinical evaluation suggested comparable interpretability between DLAC and CTAC under controlled validation conditions.