Cardiac SPECT data analysis using graph-based convolutional neural networks

Ziel/Aim:

Myocardial SPECT is a well established method to detect myocardial perfusion deficits. The analysis is typically performed on a visual basis enhanced by commercial software. A conventional approach utilizes normal databases to address gender, population and hardware specific differences in the data formation process. We investigated graph-based convolutional neural networks (GCNN) and tested both lesion detection and localization performance in a patient cohort imaged on a CZT based SPECT system.

Methodik/Methods:

Data from 450 studies with an approximate equal number of normal and abnormal studies as defined in a binary mode was included. Polar maps were used for the neural networks. We implemented 4 methods for separated analysis of rest and stress: a 1D convolutional neural network (CNN), a 2D CNN, a GCNN using Chebyshev polynomials, and a GCNN using Cayley filters. Then, localization performance was evaluated on 30 cases labeled on a segment basis by an expert using an occlusion technique. Here, one of the 17 regions is occluded and the model's probability of detect disease given the occlusion is recorded. This is repeated for each of the segments and a feature map is created. The best occlusion value was empirically determined to be a region's average amongst cases that do not present disease.

Ergebnisse/Results:

The GCNN model with Chebyshev filters achieves the highest detection rate with an accuracy of 89% and 91% respectively. However, with respect to localization, clear differences in the vessel territories where found (accuracy: RCA 95%, LAD: 64%, and LCX: 77%). The distribution of defects in ground truth images and the one from the network's abnormal detections revealed that abnormalities are higher in the RCA and LCX regions and may indicate that the network's relatively poor performance in the LAD territory may be due to the reduced number of lesions in that region.

Schlussfolgerungen/Conclusions:

The new method showed interesting results but also raised questions: as no angiographically confirmed ground truth was available, the influence of imaging hardware specific artifacts from the CZT system is open. Thus, the hard evidence and the evaluation in conventional SPECT systems is needed.