Echocardiogram-Segmentation-RandomForest

Machine learning-based segmentation of echocardiograms using Random Forest with pixel-wise feature extraction. This project segments the left ventricle in echocardiogram frames using a Random Forest classifier trained on per-pixel features including intensity, position, local statistics, and anisotropic filtering.

Results Summary

Metric	Value
Best Configuration	Trees=100, Depth=500, Criterion=Gini
Validation Dice Score	0.7120
Test Dice Score	0.6752
Test IoU (Jaccard Index)	0.5096
Test Accuracy	95.08%
Test Precision	82.27%
Test Recall (Sensitivity)	57.25%
Average Surface Distance (ASD)	4.00 pixels

Performance Interpretation

• Dice Score (0.6752): Moderate segmentation performance, clinically useful for initial screening and analysis
• High Precision (82.27%): Very few false positives, indicating the model is conservative and avoids over-segmentation
• Moderate Recall (57.25%): Some under-segmentation occurs, suggesting areas for improvement
• ASD (4.00 pixels): Average boundary error of ~4 pixels, reasonable for echocardiogram analysis

Key Features

• Pixel-wise feature extraction including:
  • Intensity values
  • Spatial coordinates (x, y positions)
  • Local statistics (mean, median, max, min, variance in 3x3 window)
  • Anisotropic diffusion filter response
• Random Forest classification with hyperparameter optimization
• Comprehensive evaluation using medical imaging metrics:
  • Dice Similarity Coefficient
  • Jaccard Index (IoU)
  • Accuracy, Precision, Recall, Specificity
  • Average Surface Distance (ASD)
• Automatic visualization of predictions

Getting Started

Prerequisites

• MATLAB R2020a or later
• Required toolboxes:
  • Image Processing Toolbox
  • Statistics and Machine Learning Toolbox

Running the Pipeline

Step 1: Extract Features

% Run this once to generate the feature table
generate_attributes('data/frames/', 'AttributesTable.mat');

Step 2: Run Segmentation

% Run the main segmentation pipeline
main_segmentation();

Configuration

The main configuration parameters in main_segmentation.m:

config.data_path = 'data/frames/';             % Path to echocardiogram frames
config.mask_path = 'data/masks/';              % Path to ground truth masks
config.attribute_file = 'AttributesTable.mat'; % Feature file name
config.num_train = 140;                        % Number of training frames
config.num_val_ratio = 0.15;                   % Validation set ratio
config.random_seed = 42;                       % For reproducibility
config.image_size = [112, 112];                % Image dimensions

Hyperparameter Tuning Results

Config	Trees	Max Depth	Criterion	Validation Dice
1	50	500	Gini	0.6969
2	50	50	Gini	0.5082
3	100	500	Gini	0.7120
4	100	50	Gini	0.5263
5	150	500	Gini	0.6842
6	50	500	Deviance	0.6719
7	50	50	Deviance	0.5462
8	100	500	Deviance	0.6922
9	100	50	Deviance	0.5033
10	150	500	Deviance	0.7110

Best Configuration: Trees=100, Depth=500, Criterion=Gini

Repository Structure

├── data/
│ ├── frames/
│ │ ├── diastole0.png ... diastole100.png
│ │ └── sistole0.png ... sistole100.png
| └── masks/
│ │ ├── diastole0.png ... diastole100.png
│ │ └── sistole0.png ... sistole100.png
├── results/                     # Results generated automatically
│ ├── results_summary.mat
│ ├── prediction_frame_54.png
│ ├── prediction_frame_56.png
│ ├── prediction_frame_155.png
│ ├── prediction_frame_192.png
│ └── prediction_frame_193.png
├── src/
│ ├── generate_attributes.m      # Feature extraction (Code 1)
│ └── main_segmentation.m        # Main pipeline (Code 2)
└── README.md                    # This file

Output Files

After running the pipeline, you'll find:

File	Description
results_summary.mat	All tested configurations and their Dice scores
prediction_frame_*.png	Individual frame visualizations

Methodology

Feature Extraction

For each pixel in every image, we extract 9 features:

1. Intensity - Original pixel value
2. X-Position - Row coordinate
3. Y-Position - Column coordinate
4. Local Mean - Mean in 3x3 neighborhood
5. Local Median - Median in 3x3 neighborhood
6. Local Max - Maximum in 3x3 neighborhood
7. Local Min - Minimum in 3x3 neighborhood
8. Local Variance - Variance in 3x3 neighborhood
9. Anisotropic Filter - Edge-preserving smoothing response

Model Training

• Algorithm: Random Forest with Bagging
• Training Data: 140 frames (70% of dataset)
• Validation: 15% for hyperparameter tuning
• Test: Remaining 15% for final evaluation

Evaluation Metrics

• Dice Coefficient: Measures overlap between predicted and ground truth
• Jaccard Index (IoU): Intersection over union
• Accuracy: Overall correct predictions
• Precision/Recall: Balance of false positives vs false negatives
• Average Surface Distance: Boundary alignment quality

License

This project is licensed under the MIT License - see the LICENSE file for details.

The dataset is provided for academic research purposes only; please cite the original authors.

Acknowledgments

• Echonet dataset for providing echocardiogram data