VeSpA Benchmarking Repository

Comprehensive benchmarking of vessel segmentation pipelines for histology images with DAB immunostaining.

Overview

This repository compares three vessel segmentation methods on histological images with brown DAB (diaminobenzidine) immunostaining marking endothelial cells:

1. VeSpA - Color-based vessel segmentation using CMYK decomposition
2. VeSpA+SAM (Hybrid) - Novel hybrid approach combining color priors with foundation models
3. SAM - Segment Anything Model with automatic mask generation
4. YOLOv8-seg - Pre-trained object detection and segmentation

Quick Start

Download the SAM model:

wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth
mkdir -p models
mv sam_vit_b_01ec64.pth models/

Install dependencies:

pip install -r requirements.txt

Run the benchmarking pipeline:

python evaluation/benchmark.py --all-categories

This will:

• Process all images in datasets/images/
• Compare all four methods
• Generate metrics (Dice, IoU, Precision, Recall)
• Create comparison plots in results/metrics/
• Save predictions in results/prediction/

---

Methods

1. VeSpA (Vessel Segmentation Pipeline Algorithm)

Approach: Color-based segmentation using CMYK decomposition

Key Steps:

• Convert BGR → CMYK color space
• Extract Yellow channel (inverse of brown stain)
• Apply Otsu threshold
• Morphological operations (blur, close, dilate, erode)

Advantages:

• Domain-specific (targets brown immunostain)
• Fast and deterministic
• No model weights required

Limitations:

• Sensitive to morphological parameter tuning
• Sometimes imperfect boundary definition

4. VeSpA+SAM (Hybrid)

Approach: Combines color priors with foundation model refinement

Key Steps:

1. Stain Detection (VeSpA): Extract brown stain using CMYK + Otsu
2. Prompt Generation: Find connected components, compute centroids
3. SAM Refinement: Use centroids as point prompts to SamPredictor
4. Post-processing: Combine masks, remove artifacts

Advantages:

• ✓ Domain-aware (targets brown stain)
• ✓ Boundary refinement (SAM correction)
• ✓ Reduced false positives (constrained to stain regions)
• ✓ Handles staining variability

Expected Improvements over SAM alone:

• Higher precision (fewer non-vessel detections)
• Better boundary accuracy
• Robustness to stain intensity variations

Expected Improvements over VeSpA alone:

• Superior boundary definition
• Less sensitive to morphological parameters
• Foundation model reliability

3. SAM (Segment Anything Model)

Approach: Foundation model with automatic mask generation

Key Steps:

• Uniform 16×16 grid of prompt points
• SAM prediction at each point
• Filtering by IoU and stability scores
• Mask combination

Advantages:

• Automatic (no prompting required)
• Detects diverse objects
• Strong generalization

Limitations:

• Not histology-aware
• Many false positives from non-vessel structures
• Grid-based sampling may miss small vessels

4. YOLOv8-seg

Approach: Pre-trained instance segmentation

Key Steps:

• Object detection + segmentation head
• Non-maximum suppression
• Mask refinement

Advantages:

• Fast inference (nano model)
• Well-optimized

Limitations:

• Trained on general objects
• May not generalize well to histology

---

Pipeline Architecture

File Structure

pipelines/
├── sam/
│   └── run_sam.py                 # Automatic SAM segmentation
├── yolo/
│   └── run_yolo.py                # YOLOv8-seg segmentation
├── VeSpA/
│   └── Segmentation + Measurements.py  # Color-based segmentation
└── hybrid/
    ├── __init__.py
    └── run_vespa_sam.py           # Novel hybrid approach (NEW)

evaluation/
├── benchmark.py                    # Main benchmarking script
└── metrics.py                      # Metric computation

datasets/
├── images/                         # Input histology images
├── groundtruth/                    # Manual annotations
└── masks/                          # Predicted masks

Configuration

All pipelines support configuration via environment variables:

# SAM configuration
export SAM_MODEL_TYPE="vit_b"          # vit_b, vit_l, vit_h
export SAM_MAX_DIM="1024"              # Input resize dimension
export SAM_CHECKPOINT="models/sam_vit_b_01ec64.pth"

# YOLO configuration
export YOLO_SEG_WEIGHTS="models/yolov8n-seg.pt"

# VeSpA/Hybrid configuration (morphological parameters)
export BLUR_KERNEL="5"
export CLOSING_KERNEL="5"

---

Evaluation Metrics

All methods evaluated on:

• Dice Coefficient: 2·|A∩B| / (|A| + |B|)
• IoU (Intersection over Union): |A∩B| / |A∪B|
• Precision: TP / (TP + FP)
• Recall: TP / (TP + FN)
• Runtime: Seconds per image

Results Output

Results saved to:

• results/metrics/results_cat2.csv - Per-image metrics (default: intersection GT)
• results/metrics/summary_cat2.csv - Mean metrics per model
• results/metrics/runtime_summary_cat2.csv - Runtime comparison
• results/metrics/results_all_categories.csv - All categories combined
• results/metrics/*.png - Comparison plots

Predictions saved to:

• results/prediction/{model_name}/*_pred.png - Binary predictions
• results/prediction/{model_name}/*_overlay.png - Overlay visualization

---

Technical Documentation

See TECHNICAL_NOTES.md for detailed explanation of:

• SAM automatic vs. prompt-guided segmentation
• Why histology-specific prompts improve segmentation
• Hybrid pipeline algorithm description
• Expected advantages and improvements

---

Citation

If you use this benchmarking repository or the hybrid VeSpA+SAM pipeline, please cite:

@misc{vespa_benchmark_2024,
  title={VeSpA Benchmarking: Hybrid Foundation Model and Color-Prior Vessel Segmentation},
  author={[Your Name]},
  year={2024},
  howpublished={GitHub},
  url={https://github.com/[your-repo]/VeSpA_benchmarking}
}

---

Requirements

• Python 3.8+
• PyTorch (CPU or CUDA)
• OpenCV
• NumPy, Pandas, Matplotlib
• segment-anything
• ultralytics (YOLO)
• scikit-image

See requirements.txt for full dependency list.

---

License

[Add appropriate license]

---

Contact

For questions or issues, please open an issue on GitHub.