Detection of Coronary Artery Disease Using Transfer Learning

Coronary artery disease (CAD) is among the most prevalent health ailments and can be life-threatening hence requiring early and accurate diagnosis for proper intervention. This paper aims to assess the performance of CNN that utilises transfer learning to differentiate abnormal perfusion in SPECTMPI. Using SPECT-MPI as data, the data used in the study were publicly shared and anonymized, and data pre-processing was done by normalizing the images, shuffling and applying geometric augmentation. Five pre-trained CNN models, including DenseNet201, ResNet101V2, InceptionV3, Xception, and EfficientNetB5, were fine-tuned using KerasTuner to perform grid search over the dropout, activation function (ReLU, ELU, Leaky ReLU, SELU), number of units per layer, and the number of dense layers. The models were trained using Adam optimizer, it uses early stopping to control for overfitting of the model. The performance was evaluated with loss, the rates of accuracy, sensitivity, and specificity, F1 score, and Fmed score. Apparently, InceptionV3 was the most successful in terms of error: 0.1440, sensitivity with 99.999%, and the specificity with 96.43%. Thus, the accuracy of the model reached 96.88% with reference to Fmed score as 0.9808, while the F1-score was equal to 0.9167. DenseNet201 and Xception also turned out to be efficient for reducing unnecessary operations and ResNet101V2 and EfficientNetB5 seem slightly less effective. Thus, the present study supports previous research in suggesting that InceptionV3's high performance might help in early CAD diagnosis and treatment in clinical practice.