Stereo 3D Reconstruction | Volume Visualisation

Project Description

As part of the Image-Guided Intervention module at Imperial College London, this project explored core computer vision methods for 3D scene understanding in medical applications. It combined stereo image analysis, geometric reasoning, and volumetric visualisation using real clinical and biological datasets.

🧠 Project Highlights

🔹 Epipolar Geometry & Tracking

• Analysed camera motion and depth perception under pure translation scenarios.
• Interpreted and illustrated epipolar constraints, explaining the impact of geometric configuration on feature correspondences.
• Proposed a real-time instrument tracking system for catheter navigation using fluoroscopy and optional electromagnetic integration.
• Developed and justified coordinate transformation pipelines for surgical navigation in the anatomical frame.

🔷 3D Scene Reconstruction 🚀

This component focused on reconstructing 3D structure from stereo endoscopic images of a surgical scene:

• Salient Features: Identified keypoints with strong geometric distinctiveness across stereo pairs to support reliable correspondence.
• Stereo Matching: Matched features between stereo views and filtered outliers based on geometric consistency.
• Fundamental Matrix Estimation: Derived the epipolar geometry governing the stereo image pair to constrain valid correspondences.
• Disparity Mapping & Depth Recovery: Generated a dense disparity map and back-projected image points to compute real-world 3D coordinates.
• Quantitative Tool Analysis: Estimated the physical diameter of a surgical tool (∼7.53 mm) directly from the reconstructed point cloud using geometric analysis, achieving 95% accuracy.

✅ Demonstrated the full stereo vision pipeline from calibrated image pairs to quantitative 3D metric extraction in a surgical context — achieving 95% accuracy in depth estimation.

🔷 Volumetric Data Visualisation 🧊

Using 3D volumetric data from optical projection tomography, this part focused on the visual interpretation of biological structure:

• Transfer Function Design: Created mappings from voxel intensity to RGBA values, enabling anatomical realism in CT renderings (bone, liver, fat).
• Maximum Intensity Projection (MIP): Implemented orthogonal MIP projections along the x, y, and z axes to highlight internal structure without segmentation.
• Interpretability: Demonstrated how projection-based rendering supports rapid insight in biomedical imaging contexts.

✅ Applied scientific visualisation techniques to enhance interpretability of 3D biomedical volumes, supporting both qualitative and quantitative analysis.

🧪 Tools & Techniques

• Python, MATLAB (SURF, epipolar geometry, disparity mapping)
• Optical Projection Tomography volume rendering
• Real-world datasets from SCARED and zebrafish imaging
• Frame transformation and registration analysis

🏫 Project Info

• Institution: Imperial College London – The Hamlyn Centre
• Programme: Image Guided Intervention for Robotic Surgery
• Supervisor: Dr. Stamatia (Matina) Giannarou
• Student: Gavin Yue
• Submitted: January 2024