A tissue-informed deep learning-based method for positron range correction in preclinical [Formula: see text]Ga PET imaging.
Authors
Affiliations (8)
Affiliations (8)
- Nuclear Physics Group and IPARCOS, Department of Structure of Matter, Thermal Physics and Electronics,, Universidad Complutense de Madrid, Av. Complutense, Pl. de las Ciencias, 1, 28040, Madrid, Spain. [email protected].
- Instituto de Investigación Sanitaria (IdISSC), Hospital Clínico San Carlos, Calle del Profesor Martín Lagos, s/n, 28040, Madrid, Spain. [email protected].
- Nuclear Physics Group and IPARCOS, Department of Structure of Matter, Thermal Physics and Electronics,, Universidad Complutense de Madrid, Av. Complutense, Pl. de las Ciencias, 1, 28040, Madrid, Spain.
- Instituto de Investigación Sanitaria (IdISSC), Hospital Clínico San Carlos, Calle del Profesor Martín Lagos, s/n, 28040, Madrid, Spain.
- Molecular Pharmacology, Memorial Sloan Kettering Cancer Center, 1275 York Avenue, New York, 10065, USA.
- Department of Radiology, Memorial Sloan Kettering Cancer Center, 1275 York Avenue, New York, 10065, USA.
- Department of Pharmacology, Weill Cornell Graduate School, 1300 York Avenue, New York, 10065, USA.
- Department of Nuclear Medicine, Bern University Hospital, Freiburgstrasse 4, 3010, Bern, Switzerland.
Abstract
Positron range (PR) limits spatial resolution and quantitative accuracy in PET imaging, particularly for high-energy positron-emitting radionuclides such as [Formula: see text]Ga. This study proposes a deep learning-based approach using 3D residual encoder-decoder convolutional neural networks (3D RED-CNNs), incorporating tissue-dependent anatomical information through a μ-map-dependent loss function. We propose a model trained from realistic simulations and, using the information in the initial PET and CT images, obtain positron range corrected images. We validated our model in both simulations and real acquisitions. Three different 3D RED-CNN architectures-Single-Channel, Two-Channel, and DualEncoder-were trained using simulated PET datasets and evaluated on both synthetic and real PET acquisitions from [Formula: see text]Ga-FH and [Formula: see text]Ga-PSMA-617 mouse studies. The performance of each model was compared to a standard Richardson-Lucy deconvolution (RL-PRC) approach using metrics such as mean absolute error (MAE), structural similarity index (SSIM), contrast recovery (CR), and contrast-to-noise ratio (CNR). In simulations, CNN-based methods achieved up to 19% improvement in SSIM and 13% reduction in MAE compared to RL-PRC. The Two-Channel model showed the highest CR and CNR values, recovering lung activity with 97% agreement to the ground truth, compared to 77% with RL-PRC. Noise remained stable for CNN models (∼5.9%), whereas RL-PRC increased noise by 5.8%. In preclinical acquisitions, the Two-Channel model achieved the highest CNR in different tissues, while maintaining the lowest noise level (9.6%). Although no ground truth was available for real data, analysis confirmed superior tumor delineation and reduced spillover artifacts with the Two-Channel model. These findings demonstrate the potential of CNN-based PRC for improving quantitative PET imaging, particularly for [Formula: see text]Ga. CNN-based methods, particularly the Two-Channel model, outperformed conventional deconvolution in both simulated and real data. Future work will focus on enhancing model generalization through domain adaptation and hybrid training strategies, as well as extending the method to other high-energy PET radionuclides.