Investigations of Scatter Correction Methods in Quantitative PET using Deep Learning

Abstract

Introduction: To render a reconstructed image acquired with PET as quantitatively accurate, corrections for the scattered coincidences have to be applied. Their contribution tends to introduce a low-frequency additive component to the acquired data, which ultimately leads to loss of contrast and erroneous SUV measurements in the reconstructed images. Despite Monte Carlo (MC)-based methods being considered the most accurate for scatter corrections, their computational demands prevent them from being clinically in use. In this work, the use of Deep Learning (DL) was investigated as a method to provide MC-grade scatter corrections within clinical timeframes.

Materials & Methods: MC simulations of two types of phantoms, 9 analytical and 24 voxelized-patients, were performed in GATE. With the resulting sinogram data of prompt coincidences and attenuation factors as input and the scattered coincidences as output, a 2D U-Net was trained. Independent network trainings were performed on 18 unique datasets. Each of them was constructed by using a different subset of simulated phantoms, a different input-output pre-processing, and a different 2D view over the same sinogram data. The performances of the trained networks were evaluated on a test dataset, which was consisted of 5 simulated phantoms excluded from the trainings, using Normalized Root Mean Squared Error (NRMSE) as a metric.

Results: The best-performing network achieved an NRMSE (mean ± standard deviation) of (4.94 ±1.88) % overall, and (3.91 ±1.21) % specifically for voxelized patient phantom cases. It was trained using the projection views of the sinogram data, with no input-output blurring. With input-output blurring applied, comparable results were obtained. Full scatter estimations of a single bed position were generated within 4.8 seconds. In 64% of the test cases, using the projection views of the sinogram data resulted in a lower mean NRMSE. The inclusion of analytical phantoms decreased the performance of the network on the voxelized-phantom tests by an NRMSE of 1% on average.

Conclusions: The feasibility of using DL for scatter estimation can be claimed. Improved accuracy is achieved by using the projection views instead of the sinogram views for trainings. Valid DL methods to generate a scatter estimation can be based on both unprocessed and blurred MC-generated training data. Which of the two constitutes the optimal strategy remains inconclusive, as their quantitative accuracy must be evaluated on the final reconstructed images.