Reconstruction-informed and multidomain deep learning for generalizable CT-free attenuation correction in SPECT...

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence in Healthcare

Background:

• Deep learning (DL) offers potential for attenuation correction (AC) in SPECT myocardial perfusion imaging (MPI) without CT-derived attenuation maps (ATMs).
• Existing DL methods for AC in SPECT MPI face limitations in accuracy and generalizability.

Purpose of the Study:

• To introduce and evaluate a novel reconstruction-informed and multidomain (RIMD) deep learning framework for AC in SPECT MPI.
• To assess the RIMD framework's performance using multi-input non-AC (NAC) images and dual-domain supervision (ATM and AC domains).

Main Methods:

• Trained a SwinUnetR model on 1058 SPECT/CT MPI scans using a 5-fold cross-validation framework.
• Employed a multi-input strategy using NAC images from three OSEM reconstruction settings.
• Utilized a combined loss function optimizing both ATM and AC domains, with evaluation on internal and external datasets.

Main Results:

• The RIMD DL framework significantly outperformed direct and indirect methods in ablation studies.
• Achieved low mean relative absolute error percentages (MRAE%) for ATMs (25.02% internal, 26.31% external) and AC SPECT images (11.72% internal, 19.31% external).
• Clinical validation showed high agreement (ICC=0.98) between DL-based AC (DLAC) and CT-based AC (CTAC) images, with comparable diagnostic accuracy.

Conclusions:

• The proposed RIMD method effectively improves indirect DL-based AC in SPECT MPI by utilizing multiple reconstruction inputs and joint loss optimization.
• The model demonstrates strong generalization on external data and robust performance in quantitative and qualitative assessments.
• Preliminary clinical evaluation suggests DLAC is a viable alternative to CTAC, offering comparable interpretability.