Recovering 3D Shape with Absolute Size from Endoscope Images Using RBF Neural Network

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Geometry

Background:

• Medical diagnosis of polyps relies on empirical assessment of 2D endoscopic images.
• Accurate 3D polyp shape and size recovery from 2D images is crucial for improved diagnostic support.
• Existing methods like the Vogel-Breuß-Weickert (VBW) model offer fast 3D shape recovery but lack exact size information.

Purpose of the Study:

• To develop a method for accurate 3D polyp shape recovery with exact size from 2D endoscopic images.
• To enhance the diagnostic capabilities in medical imaging by providing precise 3D structural information.
• To overcome the limitations of relative shape recovery in existing 3D reconstruction techniques.

Main Methods:

• Proposed a shape modification technique to refine 3D shapes recovered by the VBW model.
• Utilized Radial Basis Function Neural Networks (RBF-NN) for mapping input and output gradient parameters.
• Trained the RBF-NN using gradient parameters from VBW model outputs and true gradient parameters of generated spheres.

Main Results:

• Successfully recovered exact 3D polyp shapes with accurate size by modifying VBW model outputs.
• Demonstrated the capability of RBF-NN to learn the mapping for gradient parameter correction.
• Validated the proposed approach through computer simulations and real-world experimental data.

Conclusions:

• The novel shape modification method effectively enables precise 3D polyp size and depth recovery from 2D endoscopic images.
• The integration of RBF-NN with the VBW model significantly improves the accuracy of 3D shape reconstruction.
• This technique offers a promising tool for enhanced medical diagnosis and decision-making in endoscopy.