Improved myocardial perfusion PET imaging using artificial neural networks

Objectives: Artificial neural networks (ANNs) have proven to be powerful tools for medical image analysis and led to improvements in diagnosis and assessment of diseases. The goal of this study is to improve myocardial perfusion (MP) PET imaging in defect detection through the development of an ANN fusion scheme that integrates information from the maximum-likelihood (ML) and the post-smoothed ML reconstructions.

Methods: We developed a patch-based ANN fusion technique based on a collection of the ML and filtered ML reconstructed images for MP PET imaging. Using the XCAT phantom, we simulated three MP imaging datasets, one with normal perfusion and the other two with a non-transmural and a transmural perfusion defects. These two defects were simulated on the same region of the left ventrical with the same level of perfusion reduction. The time activity curves representing the typical patient Rb-82 bio-distribution were applied in the analytical simulation with 3 and 4.5 minutes cumulated activities. One-hundred Poisson noise fluctuations were generated for each of the three datasets at each of the two count levels. After performing the ML reconstruction, the images were post-smoothed using Butterworth filter with cutoff frequencies of 0.5 and 1 cycle/cm. We designed an ANN as a nonlinear multivariate regression function with four hidden layers: two fully-connected layers, a Hyperbolic Tangent activation layer, and a Euclidean loss layer. The ANN training images were obtained from the ML and the filtered ML reconstructed images at higher count level dataset of normal perfusion. A total number of 270K 3D patches of size 4×4×4 voxels were extracted from the training images as the inputs to feed the ANN. Using the back-propagation algorithm, the weights of the ANN were trained to minimize the Euclidean loss function calculated between the input patches and the corresponding reference image patches. Using the ANN model trained with 100K iterations, the proposed ANN fusion scheme was tested on the lower count level reconstructed images with normal perfusion, non-transmural defect and transmural defect, respectively. To quantitatively evaluate the proposed ANN fusion technique, we measured the tradeoff between the normalized mean squared error (NMSE) and the ensemble normalized standard deviation (NSD) on normal left ventrical. For the purpose of defect detection, the noise and contrast tradeoff was evaluated on the defect regions. Using the channelized Hotelling observer, we performed receiver operating characteristic analysis for the task of detecting non-transmural and transmural defects.

Results: The ANN fusion demonstrates improvement of the tradeoff between noise and bias over the filtered ML. Moreover, the ANN fused images achieve enhanced contrast while reaching comparable noise in the non-transmural and transmural defect regions. For the task of defect detection, the ANN fused image leads to a significant improvement of the area under the curve (AUC) value from 0.78±0.03 to 0.86±0.03 for the non-transmural defect and from 0.81±0.03 to 0.89±0.03 for the transmural defect. The two-tailed p-values of the comparisons between the ANN fusion and the filtered ML in the non-transmural and transmural defect detection tasks are smaller than 0.02.

Conclusions: We developed an ANN fusion scheme to integrate information from the ML image and the post-smoothed ML images with different cutoff frequencies. Using XCAT phantom simulation with normal and regionally reduced non-transmural and transmural perfusion, we applied the proposed technique and compared its performance with that of the filtered ML. The ANN fusion scheme improves the tradeoff between bias and noise in the normal perfusion as well as the noise and contrast tradeoff for the perfusion defects. Furthermore, the ANN fusion significantly improves the defect detectability of non-transmural and transmural defects, which demonstrates its promise for diagnostic and prognostic applications for MP imaging.