ECG Cardiovascular Disease Classification

Automated detection of cardiovascular diseases from 12-lead ECG images using Convolutional Neural Networks (CNNs).
This research investigates the role of lead-wise ECG segmentation in improving classification accuracy and explores the impact of individual leads in computer vision–based cardiovascular disease detection.

---

Overview

• Lead-wise ECG Processing: Extracts individual leads from full 12-lead ECG images using calibrated coordinates.
• CNN Classification: Deep learning model trained on individual lead images, with aggregated predictions for final classification.
• Comprehensive Evaluation: Includes accuracy, precision, recall, F1-score, and confusion matrices for both lead-wise and aggregated predictions.

---

Dataset

• Source: Mendeley Data Repository
• Full ECG Images: 928 (12-lead ECG scans).
• Processed Leads: 11,136 individual lead images (928 × 12).
• Image Format: Original ECGs in JPG, processed as 128×128 grayscale lead images.
• Classes (4 categories):
  • Normal: 284 ECGs → 3,408 leads
  • Abnormal Heartbeat: 233 ECGs → 2,796 leads
  • Myocardial Infarction (MI): 239 ECGs → 2,868 leads
  • History of MI: 172 ECGs → 2,064 leads

Train/Test Split:

• Training: 9,600 images (80%)
• Testing: 2,400 images (20%)

---

Preprocessing Pipeline

1. Lead Extraction:

   • Extracts 12 leads (I, II, III, aVR, aVL, aVF, V1–V6) from 3×4 ECG grid layout.
   • Coordinate-based segmentation with logical pixel mapping.

2. Enhancement & Cleaning:

   • Contrast (×1.2) and sharpness (×1.1) enhancement.
   • Gaussian blur + adaptive thresholding for noise removal.
   • Background removal via grayscale conversion and morphological operations.

3. Normalization:

   • Pixel scaling to range [0,1].
   • Standard resizing to 128×128 pixels.

---

Model Architecture

• Input: 128×128×1 (grayscale lead images).
• CNN Layers:
  • Conv2D(32, ReLU) → MaxPooling → Dropout
  • Conv2D(64, ReLU) → MaxPooling → Dropout
  • Flatten → Dense(128, ReLU) → Dense(4, Softmax)
• Optimizer: Adam
• Loss Function: Sparse Categorical Crossentropy

---

Evaluation Metrics

• Accuracy
• Precision
• Recall
• F1-score
• Confusion Matrices (per-class and aggregated)

---

Results

• The model demonstrates strong classification accuracy across the four categories.
• Lead-wise analysis highlights the impact of individual leads in detecting cardiovascular abnormalities.
• Aggregated predictions improve robustness compared to single-lead classification.

---

Dependencies

• TensorFlow / Keras – model training
• OpenCV – image processing
• PIL (Pillow) – image manipulation
• Matplotlib – plotting and visualization
• NumPy – numerical operations
• Scikit-learn – evaluation metrics

---

License

© 2025 Saqlain Ahmed. For academic research and personal use only.
Redistribution or commercial use is prohibited without prior permission.

Contact

saqlainjuna@gmail.com

Acknowledgments

• Dataset provided by Mendeley Data Repository: https://data.mendeley.com/datasets/gwbz3fsgp8/2
• Research conducted as part of Final Year Project (FYP).