Evaluating point detection performance and stability across region of interest (ROI) sizes for optimal landmark...

Accurate and stable keypoint detection is essential for reliable landmark tracking in minimally invasive and image-guided surgery. However, endoscopic imagery poses significant challenges due to soft-tissue deformation, respiratory motion, instrument interaction, specular reflections, and low-texture anatomical surfaces. While existing feature extractors such as Scale-Invariant Feature Transform (SIFT), Speeded-Up Robust Features (SURF), and Oriented FAST and Rotated BRIEF (ORB) emphasize global robustness, they do not analyze how region-of-interest (ROI) size influences local keypoint behaviour. In clinical workflows, feature detection is frequently confined to a small ROI around a landmark or telepointer projection, yet the effect of ROI configuration has not been systematically examined. This study introduces an ROI-aware feature detection framework that integrates Jerman vesselness filtering with KAZE detection to enhance feature stability under deformation and illumination variability. Three ROI sizes (80×80, 100×100, and 150×150) are evaluated across rotation, zoom-in/out, and liver-deformation datasets using inlier retention and Euclidean drift as measures of matching accuracy and temporal stability. Jerman filtering substantially increases detectable keypoints and reduces Lightness Order Error (LOE), reinforcing its effectiveness as a preprocessing stage. Statistical analysis using paired t-tests and Wilcoxon tests shows that the 100×100 ROI offers the best balance between keypoint density and stability, achieving mean accuracies of 43.28% (rotation) and 44.60% (zoom) with a stability rate of 20%. Comparable robustness is maintained under liver deformation despite severe homogeneity. Overall, this work provides the systematic quantification of ROI-size effects in deformable endoscopic scenes and establishes a principled framework for ROI optimization in dynamic surgical environments.