Objective: The exploration and the implementation of a deep learning method using a state-of-the-art convolutional neural network for the classification of polar maps represent myocardial perfusion for the detection of coronary artery disease.
Subjects and methods: In the proposed research, the dataset includes stress and rest polar maps in attenuation-corrected (AC) and non-corrected (NAC) format, counting specifically 144 normal and 170 pathological cases. Due to the small number of the dataset, the following methods were implemented: First, transfer learning was conducted using VGG16, which is applied broadly in medical industry. Furthermore, data augmentation was utilized, wherein the images are rotated and flipped for expanding the dataset. Secondly, we evaluated a custom convolutional neural network called RGB CNN, which utilizes fewer parameters and is more lightweight. In addition, we utilized the k-fold validation for evaluating variability and overall performance of the examined model.
Results: Our RGB CNN model achieved an agreement rating of 92.07% with a loss of 0.2519. The transfer learning technique (VGG16) attained 95.83% accuracy.
Conclusions: The proposed model could be an effective tool for medical classification problems, in the case of polar map data acquired from myocardial perfusion images.
Keywords: Cardiovascular diagnosis; Convolutional neural networks; Deep learning; Image classification; SPECT MPI scans.
© 2022. The Author(s) under exclusive licence to The Japanese Society of Nuclear Medicine.