Automatic reorientation algorithm for myocardial perfusion SPECT using segmentation

Abstract

Background

Cardiac reorientation is a necessary step in processing myocardial perfusion images. This task usually requires manual intervention and thus introduces intra- and inter-operator variability in the processing workflow that may lead to reduced reproducibility of the results.

Methods

A deep learning model was trained to perform segmentation of cardiac structures from SPECT images simulated from a real PET/CT dataset. Labels used for training were automatically generated in a semi-supervised fashion by using TotalSegmentator on CT images. Segmentation results from the trained model were used to calculate cardiac landmarks from which the cardiac axes were defined, and reorientation was performed. Automatic reorientation was compared against the manual reorientation defined by three expert nuclear cardiologists.

Results

The average rotation difference between cardiac axes calculated from predicted segmentations and ground-truth segmentations was ${5.3}^{°}\pm {3.1}^{°}$ on the simulated SPECT test dataset. In real SPECT images, the standard deviation of the angle difference between the automatic method and human experts was lower in all axes and operators compared to the maximum inter-operator standard deviation.

Conclusions

The proposed deep learning-based algorithm provides an automatic method to perform cardiac reorientation in myocardial perfusion SPECT images with an error range like the variability between operators and with the advantage of using objective anatomical landmarks for the definition of cardiac axes.