Improved denoising scheme using three-dimensional multi-zone convolutional neural filters in dedicated breast positron emission tomography images.

Abstract

Dedicated breast positron emission tomography (dbPET) has higher spatial resolution than whole-body PET and can detect smaller lesions. Therefore, it is expected to be useful in detecting early stage breast cancer and assessing treatment efficacy. However, dbPET images suffer leading to a relative increase in noise from reduced sensitivity. In a previous study, optimized noise reduction for each region was achieved by applying multiple convolutional neural networks (CNNs). However, CNN processing was performed in a two-dimensional (2D) slice plane, which resulted in image blurring when the image was observed from multiple directions using maximum intensity projection (MIP). In this study, we aimed to further reduce noise and improve visibility by extending multiple CNNs to the three-dimensional (3D) processing and optimizing them for each region. To train the CNN, data with acquisition times of 1 and 7 min were used as the input and teacher images, respectively. Furthermore, 3D volume data were used as the input, and the system was designed to output volume data after noise reduction processing. Quantitative evaluation of the proposed multiple 3D direction-denoising filter showed better performance than that of the 2D filter. Furthermore, the visibility of the MIP images improved. In addition, the quantitative evaluation of the maximum standardized uptake value (SUV_MAX) was conducted using a phantom; the results confirmed that the proposed noise reduction method ensured maintaining the reproducibility of SUV_MAX. These results indicate that the proposed method is effective for noise reduction in dbPET images.