Heart diseases are one of the main reasons that cause human death. The early-stage detection of heart diseases can prevent irreversible heart muscle damage or heart failure. Electrocardiogram (ECG) is one of the main heart signals that can be useful in early diagnosis because of its obvious peaks and segments. This paper focuses on using a methodology depending on deep learning for the diagnosis of the electrocardiogram records into normal (N), Supraventricular arrhythmia (SV), ST-segment changes (ST), and myocardial infarction (MYC) conditions. The continuous wavelet transform (CWT) converts the ECG signals to the time-frequency domain to compute the scalogram of the ECG signals and for the conversion of ECG signal from one dimension signal to a two-dimension image. In addition to this, a pertained model using transfer learning is applied based on Resnet50. Moreover, three main classifiers are verified to estimate the accuracy of the proposed system which are based on the Softmax, Random Forest (RF), and XGBoost classifier. An experiment is applied for the diagnosis of four main kinds of ECG records. Finally, the results based on the class-oriented schema achieved an accuracy of 98.3% based on Resnet50 with the XGBoost classifier. The comparison with the related previous work presented the excellent performance of the proposed methodology as it can be applied as a clinical application.