Toward Perfusion Defect Preservation in Deep Learning Denoising for Reduced-Dose Cardiac SPECT

In cardiac SPECT perfusion imaging deep learning (DL) denoising methods have been found to be effective for noise suppression in reduced-dose studies. However, DL training could lead to potential loss of perfusion defect signals (i.e. over-smoothing) due to high imbalance between perfusion defect and non-defect regions in training images. In this work, we investigate the feasibility of improving preservation of perfusion defects by including artificial defects in training a denoising network, in which the number and variability of perfusion defects were increased in an inexpensive fashion. In the experiments we demonstrated this approach with quarter-dose data from a set of 895 clinical acquisitions. The quantitative results indicate that the proposed approach can significantly improve perfusion defect detectability on the quarter-dose data. Compared to conventional DL training, the proposed method also achieved better perfusion defect preservation, yielding higher perfusion defect detection accuracy.