Triplet Multi-Kernel CNN for Detection of Pulmonary Diseases From Lung Sound Signals

Recent studies have demonstrated the notable success of Convolutional Neural Networks (CNNs) to detect respiratory diseases from Lung Sound (LS) signals. However, traditional Single-Kernel CNN (SK-CNN) methods frequently face limitations when depending solely on a singular kernel type to extract essential information. To overcome these limitations, we present a Multi-Kernel CNN (MK-CNN), designed specifically to capture a wider range of information from LS signals, which increases the accuracy and reliability of Pulmonary Diseases (PDs) detection. We also present a Triplet MK-CNN (TMK-CNN) model that combines the benefits of multi-kernel feature extraction with a triplet-based architecture to enhance detection performance. The effectiveness of these models was evaluated using LS data from a publicly accessible dataset provided by King Abdullah University Hospital (KAUH). Experimental results show that the MK-CNN model achieves an accuracy of 93.94%, indicating a 1.17 percentage-point improvement over the SK-CNN baseline, which is at 92.77%. The TMK-CNN model enhances classification accuracy to 97.98%, achieving a 4.04 percentage points over MK-CNN and a 5.21 percentage-point improvement compared to SK-CNN. These findings indicate the significant potential of MK-CNN and TMK-CNN architectures in enhancing automated PD identification, allowing for more reliable, user-friendly, and clinically useful diagnostic tools.