A semi-localized elastic net for surface reconstruction of objects from multislice images

Area of Science:

• Computer Vision
• Computational Neuroscience
• Optimization

Background:

• The traveling salesman problem (TSP) is a key combinatorial optimization challenge.
• Extracting object shapes from noisy images is crucial in computer vision.
• Existing neural network solutions for TSP cannot directly handle subset point selection.

Purpose of the Study:

• To adapt neural network approaches for TSP to address shape extraction from noisy image data.
• To develop a method capable of ignoring irrelevant points (noise) during shape contouring.
• To improve the robustness of neural network-based shape extraction.

Main Methods:

• A modified analog elastic net algorithm was developed, shifting focus from global to local convergence.
• This semi-localized elastic net was designed to tolerate and ignore noisy image points.
• The method was applied to extract pseudo-3D human lung outlines from MRI scans.

Main Results:

• The modified elastic net demonstrated an ability to ignore extraneous image points.
• The method showed tolerance to significant amounts of noise in the input data.
• Effective extraction of lung shapes was achieved, with an average effectiveness score of 0.06 for clean data and 0.1 for noisy data.

Conclusions:

• The semi-localized elastic net offers an effective solution for shape extraction from noisy image data, a variant of the TSP.
• This approach significantly improves upon existing methods by handling noise and selecting relevant points.
• The technique shows practical utility in medical imaging, such as lung outline extraction.