Evaluation of post-reconstruction filters applied on Methionine brain PET

초록

Aim/Introduction: Iterative reconstruction methods in nuclear medicine allows the reconstructed images to refect the observed data and the data acquisition process. However, it sufers from low count rate and intrinsic system noise. To mitigate the noise in reconstructed images and better quantitative analysis of lesions, post-reconstruction flters such as Gaussian flters are utilized. Previously, we have shown that non-local mean (NLM) flter with entropy-based regulation can be efective to prevent blurring which is prevalent in Gaussian flters. This allows better delineation of lesions or anatomical structures. On the other hand, deep learning(DL)-based image fltering is known to be efective in automated feature extraction and generation. By proper design of unsupervised learning, the DL network can learn to denoise nuclear medicine scans without external dataset. In this work, we evaluate post-reconstruction flters based on reader studies. Materials and Methods: We acquired list-mode 11C-Methionine brain PET data (20 minutes post-injection) of 50 subjects using Biograph mCT40 and mCT64 PET/CT scanner. From each list-mode data, we reconstructed with full-, half-, and quarter-time of events. For each reconstructed image, we generated 4 types of image with flters; No flter, Gaussian, NLM, and DL. For DL flter, we used a network consisted of only 3-dimensional convolutional layers and leaky rectifed linear unit activation functions. The network was trained for each subject with mean absolute error as fdelity and total variation as regularization. The scans were evaluated by two nuclear medicine physicians. The scans were graded in a 4-scale grade based on defnition of lesion, defnition of cortex, and noise in background. Results: In terms of defnition of lesion, the grades were higher in order of Gaussian, DL, NLM and no flter for full count, and DL, Gaussian, NLM, no flter in lower counts. In terms of defnition of cortex, the grades were unanimously higher in order of DL, Gaussian, NLM and no flter. For the background noise, the mean grades were higher in order of DL, Gaussian, NLM and no flter for full and quarter count, and DL, NLM, Gaussian and no flter for half count. The evaluators showed at least a fair correlation, and the highest correlation on defnition of lesions. Conclusion: The DL flter was most efective and the NLM was limited in denoising the PET scan overall. The DL flter shows promise at removing image noise while preserving important features in brain.