Deep Learning Suppresses Noise in Cardiac-Gated SPECT Without Distorting Left Ventricular Function

A 3D convolutional autoencoder reduced temporal image variability in cardiac-gated SPECT by 47.4% and improved LV wall uniformity by 20.6%, with high agreement (r>0.94) for functional measures like ejection fraction and volumes, in 50 studies.

• The deep learning network was pre-trained on 862 ungated SPECT studies and tested on 50 gated studies from the same Philips system, with external validation on 20 acquisitions from a GE StarGuide system.
• Post-denoising, Pearson correlation coefficients for left ventricular wall thickness, chamber volumes, and ejection fraction were all above 0.94 compared to pre-denoising measurements.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.