Abstract:
Colorectal cancer is one of the most common and life-threatening diseases worldwide, and
colonoscopy remains the most effective method for early detection and removal of precancerous polyps from the human colon. However, a major limitation of conventional
colonoscopy is the restricted field of view of the endoscope, which often prevents the
complete visualization of a polyp’s surface. Due to complex polyp shapes or difficult
camera angles, certain regions of the polyp may remain unobserved during the procedure, increasing the risk of misdiagnosis or incomplete removal. In this research, a 3D
reconstruction-based approach is proposed to improve the visualization and analysis of
colonic polyps using 2D endoscopic images. The overall framework begins with organizing
the dataset based on lesion and video annotations, followed by automatic frame selection
using mask-based conditions, image contrast, feature points, and depth quality scores to
select the most informative frames for reconstruction. To improve image quality, reflection
removal is applied using the EndoSRR framework. Depth maps are then estimated using
ZoeDepth, a state-of-the-art monocular depth estimation model. Using these depth maps,
dense point clouds are generated, and Region of Interest (ROI) extraction is performed
for accurate reconstruction of the polyp surface. Clean 3D meshes are constructed using the Ball Pivoting Algorithm (BPA), along with refinement and hole-filling techniques.
Additionally, the reconstructed 3D models are analyzed through various geometric and
structural features, including shape descriptors and curvature. Finally, silhouette projection and alignment validation are performed to compare the 3D reconstructed model with
the original 2D mask, ensuring accuracy and reliability of the reconstruction. This 3D
reconstruction pipeline provides enhanced visualization of colonic polyps, enabling better
inspection of previously unobserved regions and supporting further quantitative analysis
for future computer-aided diagnosis systems.
