Analysis of Lesion Signals in Reduced-Dose Pediatric Tc-DMSA Renal SPECTImages with Deep Learning Denoising

Recently we demonstrated the feasibility of applying a 3D convolutional autoencoder (CAE) network for suppressing the elevated imaging noise in reduced-dose renal studies with single photon emission computed tomography (SPECT) using 99mTc-labelled dimercaptosuccinic acid (DMSA). Built on this prior study, in this work we further investigate how the denoising performance by the 3D CAE network can impact on image features that are important for detection of cortical defects in reduced-dose DMSA renal SPECT images. For this purpose we analyzed the image contrast of a set of small to medium sized focal cortical lesions identified in clinical acquisitions. The quantitative results demonstrate that the denoising network can better preserve the image contrast of defect signals compared to conventional Gaussian filtering. Moreover, the denoising network can also reduce the image variability by 18.35% and 21.03% in defect and normal cortical regions (p-value < 0.05), respectively, in reduced-dose studies.