---

PulmoScope — AI-Assisted Lung Sound Analysis

Live Demo: https://pulmoscope.streamlit.app/

Full Manuscript, Slides, and Supporting Notebooks

The complete research materials for PulmoScope are provided below for reference, transparency, and reproducibility.

Slide Presentation

• PulmoScope Project Slides

Manuscript

• DL-FINAL-PROJECT-PULMOSCOPE.pdf

Exploratory and Model Analysis Notebooks

• EDA_Pulmoscope.ipynb — exploratory data analysis and feature inspection
• PulmoScope.ipynb — model development and experimentation

---

1. Project Overview

PulmoScope is a deep-learning–based assistive system for analyzing lung auscultation sounds and classifying respiratory conditions.
The system supports clinical screening by leveraging temporal sequence modeling for disease-level lung sound classification.

Multiple deep learning architectures were evaluated during development, including:

• Recurrent Neural Network (RNN)
• Long Short-Term Memory (LSTM)
• Pure Temporal Convolutional Network (TCN)
• Hybrid Temporal Convolutional Network with Spiking Neural Network (TCN–SNN)

Experimental results demonstrated that spatio-temporal convolution–based models (TCN and TCN–SNN) achieved superior performance compared to recurrent baselines, leading to their selection for final deployment and comparison.

Disclaimer: PulmoScope is not a medical device. It is intended solely for academic research and demonstration.

---

2. System Pipeline

Processing stages:

1. Lung sound acquisition
2. Signal preprocessing
3. Mel-spectrogram feature extraction
4. Temporal model inference
5. Prediction and interpretability

---

3. Dataset and Exploratory Analysis

PulmoScope is evaluated using the ICBHI 2017 Respiratory Sound Database, containing labeled lung sound recordings across multiple respiratory conditions.

---

4. Model Architecture

PulmoScope evaluates multiple temporal deep learning architectures for lung sound classification.
Recurrent models are included as baselines, while spatio-temporal convolution-based models are selected for final deployment due to superior performance.

Recurrent Neural Network (RNN) — Baseline

• Sequential temporal modeling
• Fully connected recurrent units
• Captures short-term temporal dependencies
• Used as an initial baseline for comparison

Long Short-Term Memory (LSTM) — Baseline

• Gated recurrent architecture
• Memory cell with input, forget, and output gates
• Designed to capture longer temporal dependencies
• Evaluated as an enhanced recurrent baseline

Pure Temporal Convolutional Network (TCN)

• Kernel sizes: 3, 5, 7
• Dilated convolutions for long-range temporal context
• Residual blocks for stable deep learning
• Attention module for temporal feature weighting

Hybrid Temporal Convolutional Network with Spiking Neural Network (TCN–SNN)

• Identical TCN backbone
• Parametric Leaky Integrate-and-Fire (LIF) spiking neuron module
• Sparse temporal activation for efficient event-driven processing
• Attention-based classifier head

Shared Classification Head

• Fully connected layers: 192 → 128 → 64
• GELU activation
• Dropout regularization
• Softmax output layer

---

5. Repository Structure

PulmoScope/
├── app.py                            # Streamlit web application entry point
├── assets/                             
│   ├── banner.png                    # Application header/banner image
│   └── figures/                      # Figures used in manuscript and README
│   ├── manuscript/                   # Complete Manuscript
│   ├── notebooks/                    # Exploratory and Model Analysis Notebooks
│   ├── sample_audio/                 # Sample audio for streamlit
│   ├── slides/                       # PPT slide PNGs for README
├── models/                           
│   ├── pure_tcn_config.json          # Pure TCN architecture configuration
│   ├── tcn_snn_config.json           # Hybrid TCN–SNN architecture configuration
│   ├── pure_tcn_weights.pth          # Trained Pure TCN model weights
│   └── tcn_snn_weights.pth           # Trained Hybrid TCN–SNN model weights
├── utils/                            
│   ├── preprocess.py                 # Audio loading and mel-spectrogram extraction
│   ├── inference.py                  # Model loading and prediction logic
│   ├── architectures.py              # TCN and TCN–SNN model definitions
│   ├── gradcam.py                    # Grad-CAM attention visualization
│   └── audio_utils.py                # Audio helper and signal utilities
├── requirements.txt                  # Python dependencies
└── README.md                         # Project documentation

---

6. Installation

Clone the Repository

git clone https://github.com/LadyJo02/PulmoScope.git
cd PulmoScope

Install Dependencies

pip install -r requirements.txt

Run the Application

streamlit run app.py

---

7. Demo Workflow (L.U.N.G Framework)

PulmoScope follows a simple user interaction flow:

• L – Load lung sound
• U – Understand sound patterns using AI
• N – Notify likely condition
• G – Guide clinical decision support

Live demo available at:
https://pulmoscope.streamlit.app/

---

8. Future Improvements

• Digital stethoscope integration
• Mobile deployment
• Noise-robust denoising models
• Larger clinical dataset expansion

---

9. Authors

PulmoScope was developed by:

• Genheylou Felisilda
• Nicole Menorias
• Kobe Marco Olaguir
• Joanna Reyda Santos

---

10. License

This project is intended solely for academic research and educational purposes. It is not approved for clinical use.

This project is open source and available under the MIT License.