Thème :
Application of advanced signal processing and deep learning techniques for the automated analysis and classification of biomedical audio signals
Présentation :
The primary research focus is on the application of advanced signal processing and deep learning techniques for the automated analysis and classification of biomedical audio signals, with a strong emphasis on respiratory health diagnostics. This work bridges the fields of computer science, electrical engineering, and medical informatics to create non-invasive, assistive diagnostic tools.
Core Research Themes:
Lung Sound Analysis & Pulmonary Disease Classification: A central pillar of this research involves developing robust methods to classify lung sounds (e.g., normal, crackles, wheezes) and diagnose pulmonary pathologies (such as COPD and pneumonia) using recordings from the ICBHI and other databases. The work systematically explores various time-frequency representations (spectrograms, scalograms, mel-spectrograms, gammatonegrams) and deep learning architectures (CNNs like VGG16, ResNet, AlexNet; Vision Transformers; BiLSTM networks) to identify the most effective combinations for accurate audio-based diagnosis.
Comparative Analysis of Deep Learning Models & Feature Extraction: The research employs a rigorous comparative methodology to evaluate the performance of state-of-the-art machine learning and deep learning algorithms. This includes extensive studies on transfer learning versus training from scratch, as well as benchmarking different neural network architectures (e.g., VGG16, ResNet-50, GoogLeNet) and traditional machine learning models (ELM, K-NN) for classification tasks.
Innovative Signal Processing for Biomedical Audio: A key contribution is the investigation and development of novel preprocessing and feature extraction pipelines tailored for physiological sounds. This includes the application of techniques like Constant-Q Transform (CQT), Variational Mode Decomposition (VMD), and MFCCs to handle challenges specific to biomedical signals, such as variable lengths and noise, particularly in lung and heart sound (PCG) signals.
Systematic Review and Scholarly Synthesis: The research is informed by and contributes to comprehensive scholarly reviews of the field, such as systematic reviews of deep learning applications in lung sound classification since 2015, ensuring the work is grounded in the latest advancements and identifies critical research gaps.
Research Impact & Trajectory: The work has evolved from foundational studies using machine learning on acoustic parameters to sophisticated deep learning systems leveraging pre-trained models and novel signal representations. The research output, published in reputable international journals and conferences (Springer, IEEE, De Gruyter), demonstrates a consistent trajectory towards creating more accurate, generalizable, and clinically relevant intelligent systems for respiratory and cardiovascular health monitoring.