End-to-End Fully Automated Cardiac Disease Diagnosis using SPECT Imaging and Deep Learning

Cardiac diseases are the leading causes of death worldwide. One of the efficient modalities to evaluate cardiac function is to use SPECT myocardial perfusion imaging (MPI), a cost-effective and non-invasive procedure. However, the examination of MPI images requires experience and is also time-consuming. Therefore, deep learning has many applications and could be helpful in this era. In this study, we intend to apply a deep learning-based algorithm for fully automated cardiac disease evaluation using SPECT images. Altogether, 196 patients were enrolled, half normal and the other half abnormal. Two images of rest and stress MPI SPECT were taken from each patient. Each image was reconstructed using the OSEM method and Butterworth filter. The inception-ResNet model was used to perform the automated diagnosis using a dual input channel of rest and stress images. Data were split into train/validation (60/10%) and test sets (30%). The performance of this model was evaluated using four parameters, including AUC, Accuracy, Sensitivity, and Specificity in test sets. The result of AUC: 0.784, Accuracy: 0.716, Sensitivity: 0.653, and Specificity: 0.788 were achieved by the proposed method. The developed model in this study could potentially assist physicians in better diagnosis.