Objectives

• Developed deep learning-based time-of-flight (DLToF) models to enhance non-ToF PET images, making them comparable to ToF PET images.
• Trained models with a range of radiotracers (18F-FDG, 18F-PSMA, 68Ga-PSMA, 68Ga-DOTATATE) for oncology, prostate, and neuroendocrine tumor PET imaging.
• Offered three levels of model strength (Low, Medium, High) to accommodate user preferences for contrast and noise.
• Demonstrated improved lesion detectability and quantification, particularly with the DLToF-H model.

Methodology

• Implemented a 3D U-Net network with residual and skip connections in PyTorch.
• Trained DLToF networks in a supervised session using mean squared error (MSE) loss function.
• Used the ADAM algorithm to update the networks' trainable parameters.
• Training data included 309 datasets from 11 sites, with 8 different tracers (75% FDG, 25% non-FDG).
• Adjusted the beta value of target ToF images to define the contrast and noise properties of each model (Low, Medium, High).

Results

• Quantitatively evaluated using 60 testing datasets (15 exams per 4 radiotracers).
• DLToF-H reduced non-ToF BSREM errors in lesion SUVmax:
  • 18F-FDG: from -39% to -6% (38 lesions).
  • 18F-PSMA: from -42% to -7% (35 lesions).
  • 68Ga-PSMA: from -34% to -4% (23 lesions).
  • 68Ga-DOTATATE: from -34% to -12% (32 lesions).
• Clinical reader study (4 readers) showed DLToF-H improved lesion detectability, DLToF-L had highest image quality scores, and DLToF-M had best diagnostic confidence scores.
• Liver noise measurements showed DLToF-L had lower noise than non-ToF BSREM, and DLToF-H had noise levels similar to or slightly higher than ToF images.

Discussions

• The study demonstrates a well-designed deep learning approach for enhancing non-ToF PET images. However, the testing sets were chosen with small and low-contrast lesions, potentially biasing the results towards emphasizing the gap between ToF and non-ToF reconstructions.
• The readers were shown all 5 image series of a subject simultaneously, which might introduce bias, although it could help in identifying false positives or missing lesions.
• A data sufficiency experiment for the proportion of multi-tracer datasets was not performed. Future work should address this.
• The study lacks quantitative and clinical evaluation of the trained models on Omni Legend™ PET/CT scanners. This should be a focus of future research.

---

Reference: Deep learning based time of flight ToF enhancement of non ToF PET scans for different radiotracers