Establishing a generalizable and universal fine-tuning framework for pre-trained vision networks in the ECG classification for Cardiovascular Disease(CVD) is crucial for advancing reliability and effectiveness in medical imaging. This paper proposes deep feature extraction, optimization, and uncertainty analysis along with transfer learning approaches for categorization of ECG images with pre-trained convolutional neural networks (CNN). On a labeled dataset of 1,376 ECG images split into four diagnostic classes, five well-known architectures, ResNet50, EfficientNetB0, DenseNet121, VGG16, and MobileNetV2, were systematically assessed, implying uncertainty quantification for enhanced explainability and calibration, and the framework was further analyzed on several datasets. The two-stage approach used in the classification pipeline is the first feature extraction using frozen base layers, followed by the selective fine-tuning of the deeper layers. The confusion matrix and the ROC-AUC studies support ResNet50’s exceptional classification performance among all assessed models, attesting to a validation accuracy of 99.05% and a balanced accuracy of 0.99 with uncertainty quantification (Mean Total Uncertainty: 0.0612, Mean Epistemic Uncertainty: 0.0090, Mean Aleatoric Uncertainty: 0.0523, Mean Confidence: 0.9867, Calibration Error: 0.0064, Uncertainty Separation: 0.6499). Especially in situations limited by limited annotated data, the results highlight the efficiency of using pre-trained visual networks for medical picture categorization. Using deep transfer learning, this framework lays the groundwork for scalable, high-precision ECG-based diagnostic systems.