AMMAR ALQAHTANI

Artificial Intelligence Specialization · Cohort 2026

OCT-Based Pose Estimation of Surgical Instruments

This project develops a deep-learning pipeline that estimates the articulated pose of surgical instruments directly from Optical Coherence Tomography (OCT) volumes, enabling fast, image-based feedback for robotic microsurgery. Each OCT acquisition is represented as a sparse 3D point cloud and processed by VoxelNeXt, a fully sparse convolutional network that localizes eight 3D keypoints of a four-link tool model (base, shaft, and two jaws); from these keypoints the three tip joint angles the surgeon controls (yaw, pitch, and jaw opening) are derived analytically. Ground truth is produced with an interactive 3D annotation tool accelerated by a model-assisted human-in-the-loop scheme, and labels are stored in physical millimetres so that annotations made on high-resolution scans transfer to the low-resolution previews used at run time. The same pipeline was shown to generalize across two different instruments, the da Vinci Research Kit (dVRK) tool and a pair of surgical scissors, localizing keypoints with sub-millimetre accuracy and recovering the tip angles to a few degrees on held-out recordings.

Read the final report (PDF)