Deep learning‑based CT-free attenuation and scatter correction for pediatric whole-body PET imaging (2024)

Methods: We retrospectively processed data of 83 whole-body PET/CT scans from 58 pediatric patients under 12 years of age (6.1 ± 0.6 years, weight 23.1 ± 12.9 kg) without gross misregistration between PET and CT. All patients underwent ¹⁸F-FDG (0.11 ± 0.01 mCi/kg administered activity) examinations on a GE Discovery MI Gen 2 PET/CT scanner using a standard-of-care acquisition protocol. For each scan, the PET data were reconstructed using the manufacturer standard reconstruction (Q.Clear, regularization parameter = 700) with no ASC (NASC) and with ASC based on the CT generated μ-map, the latter serving as the reference for evaluation.

We developed two DL methods, with method A training a CNN to predict an ASC PET image (λ-CNN PET) directly from the NASC PET image out of which method B predicts a μ-map. In method B, to focus the CNN on predicting a μ-map corresponding to the patient body, we divided the CT-generated μ-map into two parts: patient body and everything (equipment/bedding) outside of the body. The CNN predicted only the μ-map for the patient body and objects outside of the body were added back to the CNN-generated pseudo μ-map before Q.Clear reconstruction was performed (μ-CNN PET). Out of the 83 scans, data from 63 were used for training and 20 were used for performance analysis. Random rotation of the training input and label was used to increase the robustness of the model and prevent overfitting. Ninety-one thousand training data pairs were generated for each model which used an ADAM optimizer with a learning rate of 0.0001 and L2 loss. To quantitatively evaluate the λ-CNN PET and μ-CNN PET results against the CT-based ASC PET as the reference, standardized uptake values (SUVs) were measured in the liver (3-cm diameter region of interest [ROI]) and in up to five of the most-avid lesions (if lesions were present). Twenty-seven total lesions from 15 scans were included. Dunn’s test was used to compare SUV_mean and SUV_max between reconstruction results with p < 0.05 considered significant. The structural similarity index measure (SSIM) was employed to compare the similarity between λ-CNN PET or μ-CNN PET and the reference CT-based ASC PET.

Conclusions: In conclusion, for pediatric patients imaged on a state-of-the-art digital PET/CT system, ASC can be achieved using a CNN-predicted attenuation map without the use of CT. Resultant PET images show no significantly difference from CT-based ASC PET images in SUV measures.