Noise Reduction and Categorization of Cardiac Audio Images via Integrated Deep Learning Technique

Cardiovascular ailments remain a leading contributor to global mortality, making early and accurate detection vital. Early detection and management significantly reduce mortality rates. Conventional methods such as ECG, echocardiography, and MRI have drawbacks like high cost and operational complexity. This research introduces a novel approach for classifying heart conditions based on Phonocardiogram (PCG) signals, which are easier to acquire and cost-efficient. Initially, MATLAB-based denoising techniques eliminate background noise from PCG signals. Subsequently, Continuous Wavelet Transform (CWT) translates the one-dimensional PCG signals into images, capturing vital time-frequency details. The integrated model combines VGG16's capability for capturing detailed spatial characteristics and InceptionV3's multi-scale feature extraction proficiency. Extensive data augmentation enhances the model's robustness, achieving a validation accuracy of approximately 89.58 %, demonstrating strong resistance to Additive White Gaussian Noise (AWGN). Comparatively, the proposed model surpasses noise-adapted classifiers such as MobileNet and NRC-Net. This methodology presents a reliable, cost-effective, non-invasive solution for early cardiac disease detection, thereby potentially enhancing clinical outcomes through precise diagnostics.