Accelerated SPECT image reconstruction with FBP and an image enhancement convolutional neural network

Abstract

BACKGROUND: Monte Carlo-based iterative reconstruction to correct for photon scatter and collimator effects has been proven to be superior over analytical correction schemes in single-photon emission computed tomography (SPECT/CT), but it is currently not commonly used in daily clinical practice due to the long associated reconstruction times. We propose to use a convolutional neural network (CNN) to upgrade fast filtered back projection (FBP) image quality so that reconstructions comparable in quality to the Monte Carlo-based reconstruction can be obtained within seconds. RESULTS: A total of 128 technetium-99m macroaggregated albumin pre-treatment SPECT/CT scans used to guide hepatic radioembolization were available. Four reconstruction methods were compared: FBP, clinical reconstruction, Monte Carlo-based reconstruction, and the neural network approach. The CNN generated reconstructions in 5 sec, whereas clinical reconstruction took 5 min and the Monte Carlo-based reconstruction took 19 min. The mean squared error of the neural network approach in the validation set was between that of the Monte Carlo-based and clinical reconstruction, and the lung shunting fraction difference was lower than 2 percent point. A phantom experiment showed that quantitative measures required in radioembolization were accurately retrieved from the CNN-generated reconstructions. CONCLUSIONS: FBP with an image enhancement neural network provides SPECT reconstructions with quality close to that obtained with Monte Carlo-based reconstruction within seconds.