Evaluation of methods for generative modeling of cell and nuclear shape

Area of Science:

• Quantitative biology
• Biophysics
• Computational imaging

Background:

• Cell shape is crucial for function, yet robust generative modeling remains challenging.
• Existing methods for cell shape analysis lack comprehensive evaluation for generative model accuracy.

Purpose of the Study:

• To compare traditional and deep learning generative models for cell shape.
• To evaluate reconstruction accuracy across various cell image datasets (2D and 3D).
• To develop and assess improved methods for complex 3D cell and nuclear shape modeling and dynamics.

Main Methods:

• Comparison of traditional methods (outline-based) and deep autoencoders (image-based).
• Development of an improved spherical harmonic transform method for 3D shapes.
• Evaluation of shape dynamics via interpolation in shape space.

Main Results:

• Outline-based methods outperformed image-based autoencoders for shape reconstruction.
• Deep autoencoders produced overly smooth shapes, even with high-dimensional encodings.
• The improved spherical harmonic method significantly enhanced 3D shape reconstruction and reduced deformation energy during interpolation.
• Similar performance was observed for joint cell and nuclear shape modeling.

Conclusions:

• Improved spherical harmonic methods provide superior cell and nuclear shape modeling.
• These methods offer better representations, computational efficiency, and require fewer training images than deep learning approaches.
• The developed methods facilitate practical shape evolution in high-dimensional spaces.