Riemannian Manifolds for Biological Imaging Applications Based on Unsupervised Learning.

Area of Science:

• Computational Biology
• Biotechnology
• Medical Imaging

Background:

• Neural networks and multimodal systems are increasingly important in biological research.
• Computer vision techniques, particularly for cell image analysis, are underutilized.
• Unsupervised learning for image clustering, especially for single cells, presents significant opportunities.

Purpose of the Study:

• To evaluate the feasibility of using latent representations for single-cell clustering in medicine and biotechnology.
• To explore the application of embeddings for morphological cell characterization.
• To investigate neural network applications in muscle differentiation studies using C2C12 cells.

Main Methods:

• Utilizing latent space analysis for extracting morphological features from cell images.
• Applying unsupervised learning techniques, including graph-based methods and hyperbolic embeddings.
• Investigating clustering and segmentation in non-Euclidean spaces using the Poincaré ball model.

Main Results:

• Demonstrated the potential of latent representations for morphological feature extraction in cell images.
• Showcased the applicability of unsupervised learning for single-cell image clustering.
• Highlighted the value of hyperbolic embeddings for vision tasks like clustering.

Conclusions:

• High-quality latent representations from phase-contrast or bright-field images open new avenues for visual-language models in biology.
• Unsupervised segmentation in non-Euclidean spaces is a promising, yet undervalued, research area.
• This work lays the foundation for integrating visual and biological data into multimodal spaces for in vitro studies.