Abstract
1541
Introduction: Myocardial SPECT is a widely used non-invasive method to detect coronary artery disease and data analysis is performed on a visual basis supported by highly automated software. The latter is typically based on a comparison to so-called normal databases addressing gender, population and hardware specific differences. We sought to improve these approaches using Graph-based Convolutional Neural Networks (GCNN) and evaluated both disease detection and localization performance in a patient cohort imaged on a CZT SPECT system.
Methods: We included data from 503 rest and 443 stress studies with an approximately equal number of normal and abnormal studies as defined by expert review. Image data was processed using volumetric data analysis. The resulting polar maps were then used for the neural networks with 135 elements. Disease detection was performed using the 17 segment model. We evaluated 4 neural network architectures for separated analysis of rest and stress cases: a 1D fully connected neural network (CNN), a 2D CNN, a GCNN using Chebyshev polynomials (1), and a GCNN using Cayley filters. Localization performance was evaluated with the best model on 30 polar maps labeled on a segment basis by an expert reader using an occlusion technique following the work of Zeiler (3). 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 to obtain a heat map showing the importance for each segment for the overall decision. This enables us to see which areas of the polar map the model is most sensitive to, i.e. where the abnormality potentially is. Occlusion was performed by assigning the same value to each node of a particular region while maintain the rest of the graph unchanged. The best occlusion value was empirically determined to be a region’s average amongst cases that do not present CAD. The networks were trained on a system with 128GB RAM, Intel(R) Xeon(R) @ 3.50GHz and a 12 GB GeForce GTX TITAN X graphics card. Graph-based models were implemented with TensorFlow. Baseline methods were implemented with Keras. All model have a learning rate of 1e-3 and are optimised with Stochastic Gradient Descent (SGD). Baseline methods are trained in mini batches of 64 samples for 70 epochs. Both graph-based methods were run for 60 epochs, with the Chebyshev model using a mini batches of 30 samples and the Cayley model with a mini batches of 20. RESULTS: The proposed GCNN model using Chebyshev filters achieves the highest performance of disease detection with an accuracy of 89% and 91% respectively. However, with respect to localization, differences between the vessel territories where found (accuracy: RCA 95%, LAD: 64%, and LCX: 77%). Comparing the distribution of abnormalities in ground truth images to the one obtained from the network’s abnormal detections revealed that abnormalities are more common in RCA and LCX regions and may indicate that the network’s poor performance in the LAD territory may be due to the reduced number of cases in the training set exhibiting abnormalities in that region.
Conclusions: The proposed method showed promising results but also raised questions: In the absence of angiographically confirmed ground truth, the question remains unanswered whether this is influenced by artifacts from the CZT system. Thus, we aim at refinements in a) a cohort with homogeneously distributed disease and b) the evaluation in conventional SPECT systems. REFERENCES: (1) Defferrard, M., et al. Convolutional neural networks on graphs with fast localized spectral filtering. In Advances in Neural Information Processing Systems, 3844-3852 (2016).(2) Levie, R., Monti, F., Bresson, X. ,Bronstein, M. M. Cayleynets: Graph convolutional neural networks with complex rational spectral filters. arXiv preprint arXiv:1705.07664 (2017).(3) Zeiler, M. D. Fergus, R. Visualizing and understanding convolutional networks. In Europ. conference on computer vision, 818-833 (2014).