Objectives

Developed deep learning (DL) models for CT-free attenuation correction (AC) and Monte Carlo-based scatter correction (SC) in quantitative 90Y SPECT imaging.
Trained three separate models for AC, SC, and joint attenuation and scatter correction (ASC) using a modified 3D Swin UNETR architecture.
Achieved fast inference times (20 seconds per SPECT image on GPU, 6 minutes on CPU) compared to traditional methods (80 minutes on CPU).
Used dose maps, derived from corrected and uncorrected SPECT images, as input and output for the models, leveraging meaningful dose values (Gy).
Evaluated model performance extensively at both organ and voxel levels, demonstrating comparable accuracy to a commercial reconstruction tool.

Methodology

Utilized a modified 3D Swin UNETR architecture, combining a U-shaped network with a Swin Transformer encoder and a Convolutional Neural Network (CNN) decoder.
Employed a 5-fold cross-validation training scheme on a dataset of 190 patients who underwent 90Y selective internal radiation therapy (SIRT).
Used random patch sampling (64x64x64 voxels), rotation, and flipping for data augmentation, with dropout set to 0.1.
Trained models using the Adam optimizer and L1 loss function (Mean Absolute Error) for 200 epochs.
Performed test image inference using a sliding window method with the same patch size as training.

Voxel-level quantitative metrics (average±SD) for the AC task: mean error (ME (Gy)): -0.026±0.06, structural similarity index (SSIM (%)): 99.5±0.25, peak signal to noise ratio (PSNR (dB)): 47.28±3.31.
For the SC task: ME: −0.014±0.05, SSIM: 99.88±0.099, PSNR: 55.9±4.
For the ASC task: ME: -0.04±0.06, SSIM: 99.57±0.33, PSNR: 47.97±3.6.
Voxel-level gamma evaluations with criteria “DTA: 4.79, DD: 1%”, “DTA:10 mm, DD: 5%”, and “DTA: 15 mm, DD:10%” yielded pass rates around 98%.
Mean absolute error (MAE (Gy)) for tumor and whole normal liver (WNL): AC (7.22±5.9 and 1.09±0.86), SC (8±9.3 and 0.9±0.8), and ASC (11.8±12.02 and 1.3±0.98).

The study utilizes a relatively large dataset (190 patients), but it is from a single center, potentially limiting generalizability. Future work should include multi-center data.
The models' performance is affected by the presence of metal implants and heterogeneous tissue compositions (e.g., calcifications, air-tissue interfaces). The authors should consider incorporating strategies to address these challenges, such as including more diverse training data with these features or developing specific modules to handle them.
While the study compares the DL models to a commercial reconstruction tool, a more direct comparison with other published deep learning-based methods for AC and SC in 90Y SPECT would strengthen the findings.
The rationale for using dose maps instead of SPECT images directly is well-justified. However, providing more details on the limitations encountered when training with SPECT images (supplementary material) would be beneficial.