Medical Image Analysis - Vessel Segmentation and Filtering

This project, completed as part of the Image-Guided Intervention module at Imperial College London, explores classical image processing techniques for analyzing retinal images from the DRIVE dataset. The goal is to detect and segment blood vessels using low-level computer vision methods without relying on deep learning.

Figure: Retinal image used for vessel segmentation

1. Edge Detection
Various 3×3 kernels were implemented to detect horizontal, vertical, and diagonal lines. The Laplacian of Gaussian (LoG) operator was applied with different kernel sizes to examine sensitivity and noise robustness across scales.

Figure 1: Edge detection based on different sizes of LoG kernel

To summarise, a smaller kernel size heightens sensitivity to finer edges within the image (e.g. KZ = 7). This sensitivity, however, may also intensify noise, resulting in a noisier edge map (KZ = 5). When the kernel size is too small concerning the σ value, severe truncation can occur (KZ =3). On the contrary, larger kernel sizes capture broader edges and are more robust to noise but might overlook finer image details.

2. Image Filtering
Using the Fourier Transform, high-pass filters were applied in the frequency domain to enhance edges and details. Gaussian smoothing was explored both in spatial and frequency domains. The impact of different σ values on image blurring and detail preservation was visualized and compared.

• First step: Convert the 2D image to the frequency domain Fourier transform converts a function into an alternative representation. In this case, it converts the spatial information of the image into its frequency representation.
• Second step: Design and apply the high-pass filter Create a filter to eliminate low frequencies and let high frequencies pass to enhance it (e.g. the edge-detection filter). Then multiply the frequency spectrum by the high-pass filter.
• Third step: Inverse Fourier transform for reconstruction Convert the modified spectrum back to the spatial domain of the image.

Figure 2: Apply image smoothing using 2D convolution with a gaussian kernel of standard deviation 𝜎 = 1 and size [(2 ∗ 𝑐𝑒𝑖𝑙(2 ∗ 𝜎) + 1) × (2 ∗ 𝑐𝑒𝑖𝑙(2 ∗ 𝜎) + 1)]

3. Image Segmentation
Threshold-based segmentation was used to extract retinal blood vessels. Accuracy was evaluated against ground truth using ROC analysis. The optimal threshold was identified (≈0.101) based on minimum distance to (0,1) on the ROC curve. Results demonstrated how threshold choice affects vessel detection accuracy and noise.

Figure 3: Segmentation accuracy and anaylsis

📊 Tools & Dataset

• Dataset: DRIVE Retinal Image Dataset
• Techniques: LoG Edge Detection, Gaussian Smoothing, Frequency Domain Filtering, ROC Curve Evaluation
• Languages: Python (NumPy, OpenCV, Matplotlib), MATLAB

🎓 Proejct Info

• Institution: Imperial College London – The Hamlyn Centre
• Project Supervisor: Dr. Stamatia (Matina) Giannarou
• Student: Gavin Yue
• Submission: November 2023

This project served as a foundation in medical image processing, reinforcing key concepts in signal filtering, segmentation logic, and performance evaluation, essential for advancing toward AI-powered solutions in healthcare imaging.