PROJECT TITLE :

Use of a Tracer-Specific Deep Artificial Neural Net to Denoise Dynamic PET Images

ABSTRACT:

The use of kinetic modeling (KM) on a voxel level in dynamic PET pictures frequently results in large amounts of noise, lowering the precision of parametric image analysis significantly. In this research, we look at how to denoise dynamic PET scans using machine learning and artificial neural networks. We used noisy and noise-free spatiotemporal image patches recovered from simulated images of [ 11 C]raclopride, a dopamine D2 receptor agonist, to train a deep denoising autoencoder (DAE). Traditional denoising techniques, such as temporal and spatial Gaussian smoothing, iterative spatiotemporal smoothing/deconvolution, and highly limited backprojection processing, are compared to the DAE-processed dynamic and parametric pictures (simulated and recorded) (HYPR). With temporal and 5.90 percent (14.50 percent) with spatial smoothing, 5.82 percent (16.21 percent) with smoothing/deconvolution, 5.49 percent (13.38 percent) with HYPR, and 3.52 percent (11.41 percent) with DAE, the simulated (acquired) parametric image non-uniformity was 7.75 percent (19.49 percent). In terms of the coefficient of variation of voxel values and the structural similarity index, the DAE delivered the best results. With temporal and 26.31 percent with spatial smoothing, 28.61 percent with smoothing/deconvolution, 27.63 percent with HYPR, and 14.8 percent with DAE, denoising-induced bias in the regional mean binding potential was 7.8% and 26.31 percent, respectively. Erroneous results were achieved when the test data did not match the training data. Our findings show that, as compared to traditional spatiotemporal denoising approaches, a deep DAE may significantly reduce voxel-level noise while introducing a similar or lower amount of bias. The higher DAE performance comes at the expense of lower generality and the need for more training data.