Deep learning-guided attenuation and scatter correction of <sup>99m</sup>Tc-MAA SPECT images: towards quantitative analysis in <sup>90</sup>Y-SIRT.
Authors
Affiliations (7)
Affiliations (7)
- Division of Nuclear Medicine and Molecular Imaging, Geneva University Hospital, Geneva, CH-1211, Switzerland.
- Centre for Biomedical Imaging (CIBM), Geneva, Switzerland.
- Laboratory of Neuroimaging and Innovative Molecular Tracers (Nimtlab), Geneva University Neurocenter and Faculty of Medicine, University of Geneva, Geneva, Switzerland.
- Division of Nuclear Medicine and Molecular Imaging, Geneva University Hospital, Geneva, CH-1211, Switzerland. [email protected].
- Department of Nuclear Medicine and Molecular Imaging, University of Groningen, University Medical Center Groningen, Groningen, Netherlands. [email protected].
- Department of Nuclear Medicine, University of Southern Denmark, Odense, Denmark. [email protected].
- University Research and Innovation Center, Óbuda University, Budapest, Hungary. [email protected].
Abstract
This study aimed to develop deep learning (DL) models for CT-free attenuation correction and Monte Carlo-based scatter correction in <sup>99m</sup>Tc-macroagregated albumin (<sup>99m</sup>Tc-MAA) SPECT imaging, with the goal of enhancing quantitative accuracy for improved treatment planning and pre-therapy dosimetry in <sup>90</sup>Y-selctive internal radiation therapy (SIRT). Data from 222 patients who underwent <sup>99m</sup>Tc-MAA SPECT imaging prior to <sup>90</sup>Y-SIRT were included in this study. Uncorrected SPECT images (without attenuation and/or scatter correction) were used as input to a modified 3D shifted-window UNet Transformer (Swin UNETR) architecture. Three separate models were trained to predict attenuation corrected (AC), scatter corrected (SC), and joint attenuation and scatter corrected (ASC) SPECT images. The dataset was split into a training set (~ 80%) and an independent test set (~ 20%). Model training was performed using a five-fold cross-validation framework, with final evaluation conducted on the blind test set. To clinically assess model performance, 3D voxel-wise dosimetry was performed on the test set using the local energy deposition method, assuming <sup>99m</sup>Tc-MAA as a surrogate for <sup>90</sup>Y distribution. Quantitative evaluation included organ- and voxel-level metrics, along with Gamma analysis using three combinations of distance-to-agreement (DTA, mm) and dose-difference (DD, %) criteria. The average (± SD) of the voxel-wise mean error (ME) was ≤ 0.003 Gy for all tasks. The Relative Error (RE (%)) for AC, SC, and ASC tasks were 4.64 ± 7.52%, 8.99 ± 26.35%, and 16.45 ± 25.83%, respectively. Voxel-level Gamma evaluations within the whole body using three different criteria sets, including "DTA: 4.79 mm, DD: 1%"; "DTA: 10 mm, DD: 5%"; and "DTA: 15 mm, DD: 10%" yielded pass rates of over 99.60%. The mean absolute error (MAE) for lesions, normal liver and lungs across all tasks were 3.16 ± 3.39, 0.35 ± 0.36, 0.41 ± 0.47 Gy for AC, 1.97 ± 2.79, 0.19 ± 0.16, 0.22 ± 0.20 Gy, for SC and 5.16 ± 7.10, 0.45 ± 0.51, and 0.34 ± 0.37 Gy for ASC, respectively. Multiple models were developed for key SPECT quantification tasks, with potential value in clinical setting lacking reliable CT data or sufficient computational resources for Monte Carlo simulations. The models look promising for potential clinical translation and integration into commercial reconstruction software.