A tissue-informed deep learning-based method for positron range correction in preclinical 68Ga PET imaging

TL;DR

Addressing PR blurring in $^{68}$Ga PET, the paper introduces a tissue-informed deep learning framework using 3D RED-CNNs with a μ-map–guided, mutual-information–based loss to enable regionally adaptive PRC. Three architectures (Single-Channel, Two-Channel, DualEncoder) are evaluated against a Richardson-Lucy baseline, with training on synthetic MOBY/Digimouse phantoms and validation on simulated data and two real mouse acquisitions ($^{68}$Ga-FH and $^{68}$Ga-PSMA-617). The Two-Channel model consistently yields the best quantitative and qualitative results, achieving higher CR/CNR and stable noise, while real-data correction improves tumor delineation and reduces spillover. The study demonstrates the potential of tissue-informed DL-based PRC to enhance quantitative PET imaging with $^{68}$Ga and outlines domain adaptation strategies to improve generalization to real-world data.

Abstract

Positron range (PR) limits spatial resolution and quantitative accuracy in PET imaging, particularly for high-energy positron-emitting radionuclides like 68Ga. We propose a deep learning method using 3D residual encoder-decoder convolutional neural networks (3D RED-CNNs), incorporating tissue-dependent anatomical information through a u-map-dependent loss function. Models were trained with realistic simulations and, using initial PET and CT data, generated positron range corrected images. We validated the models in simulations and real acquisitions. Three 3D RED-CNN architectures, Single-channel, Two-channel, and DualEncoder, were trained on simulated PET datasets and evaluated on synthetic and real PET acquisitions from 68Ga-FH and 68Ga-PSMA-617 mouse studies. Performance was compared to a standard Richardson-Lucy-based positron range correction (RL-PRC) method using metrics such as mean absolute error (MAE), structural similarity index (SSIM), contrast recovery (CR), and contrast-to-noise ratio (CNR). CNN-based methods achieved up to 19 percent SSIM improvement and 13 percent MAE reduction compared to RL-PRC. The Two-Channel model achieved the highest CR and CNR, recovering lung activity with 97 percent agreement to ground truth versus 77 percent for RL-PRC. Noise levels remained stable for CNN models (approximately 5.9 percent), while RL-PRC increased noise by 5.8 percent. In preclinical acquisitions, the Two-Channel model achieved the highest CNR across tissues while maintaining the lowest noise level (9.6 percent). Although no ground truth was available for real data, tumor delineation and spillover artifacts improved with the Two-Channel model. These findings highlight the potential of CNN-based PRC to enhance quantitative PET imaging, particularly for 68Ga. Future work will improve model generalization through domain adaptation and hybrid training strategies.

Figures (6)

• Figure 1: 3D RED-CNN architectures implemented: A shows the Single-Channel 3D RED-CNN, which processes only PET input; B represent the Two-Channel 3D RED-CNN, where $\mu$-map information is added as a second channel; and C, the DualEncoder 3D RED-CNN, where $\mu$-map information is integrated into the network through a dedicated encoder. The legend in C indicates the network components: blue boxes represent 3×3×3 convolutional layers, orange boxes represent 3×3×3 deconvolutional layers, gray boxes denote ReLU activation functions, and connecting symbols skip connections with summation. The input images are previously normalized.
• Figure 2: Overview of the data preprocessing and sampling workflow. Each 3D image ($^{18}$F, $^{68}$Ga, and $\mu$-map) was first clipped at the 99.99995th percentile to remove hot pixels. A normalization factor was then computed for each volume and stored without modifying the raw data. All datasets and their corresponding normalization factors were saved in an HDF5 file, grouped by numerical phantom (Digimouse or MOBY). During training, the DataLoader and custom Sampler dynamically read patches from the HDF5 file, apply the stored normalization factor, and perform random 3D flips before feeding them into the network.
• Figure 3: A shows axial and coronal PET images comparing ground truth $^{18}$F with $^{68}$Ga input, $^{68}$Ga PET image PRC with Richardson-Lucy deconvolution and the output of the three models (Single-Channel, Two-Channel and DualEncoder 3D RED-CNN). The reconstructions aim to replicate the $^{18}$F ground truth distribution using $^{68}$Ga tracers through different correction methods. The white letters indicate the image direction: A=anterior, L=left, R=right. B shows the myocardium line profile in the axial view at 0$^{\circ}$ and C the bladder line profile in the coronal view at 0$^{\circ}$ . Abbreviations: GT=Ground Truth, RL-PRC=Richardson-Lucy Positron Range Correction. Results are given in Standardized Uptake Values relative (SUVr) to the mean activity in the mouse.
• Figure 4: Evaluation of image quality using SSIM and MAE for $^{68}$Ga images without and with PRC: A and B show violin plots displaying SSIM and MAE values for 20 samples. C and D show a slice coronal MAE and mean SUV difference maps of a single case. Abbreviations: RL-PRC=Richardson-Lucy Positron Range Correction, SSIM=Structural Similarity Index Measure, MAE=Mean Absolute Error.
• Figure 5: A CR ($\%$) and B CNR comparison across organs -myocardium, bladder, lungs and liver- and methods. Abbreviations: RL-PRC=Richardson-Lucy Positron Range Correction, CR=Contrast Recovery, CNR=Contrast-to-Noise Ratio,